diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml index 620cab950dc..21a3ddf07fb 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049, Device 0050] @@ -208,78 +208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml index 0d45b44282a..c684943cfdc 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049, Device 0050] @@ -212,79 +212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml index 03ca6fdffee..06e0a6550b8 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049, Device 0050] @@ -184,74 +184,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_BBS_BH_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_BBS_BH_UserArgs.yaml index c52fe8123b1..3fce135a746 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_BBS_BH_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_BBS_BH_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -97,7 +97,7 @@ ClusterLocalRead: 1 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_BBS_BH_MT256x256x64_MI16x16x1_UserArgs_shortname0_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_BBS_BH_MT256x256x64_MI16x16x1_UserArgs_shortname0_gfx950 DebugStreamK: 0 DepthU: 64 DirectToLds: true @@ -233,83 +233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -399,7 +322,7 @@ ClusterLocalRead: 1 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_BBS_BH_MT256x256x64_MI16x16x1_UserArgs_shortname1_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_BBS_BH_MT256x256x64_MI16x16x1_UserArgs_shortname1_gfx950 DebugStreamK: 0 DepthU: 64 DirectToLds: true @@ -535,83 +458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -701,7 +547,7 @@ ClusterLocalRead: 1 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_BBS_BH_MT256x256x64_MI16x16x1_UserArgs_shortname1_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_BBS_BH_MT256x256x64_MI16x16x1_UserArgs_shortname1_gfx950 DebugStreamK: 0 DepthU: 64 DirectToLds: true @@ -837,83 +683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -998,4 +767,4 @@ - null - null - DeviceEfficiency -- Equality \ No newline at end of file +- Equality diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_F8B8BS_BH_SAB_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_F8B8BS_BH_SAB_UserArgs.yaml index 01d57dcb1fd..4babe78dc32 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_F8B8BS_BH_SAB_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_F8B8BS_BH_SAB_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -97,7 +97,7 @@ ClusterLocalRead: 0 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_F8B8BS_BH_SAB_UserArgs_shortname0_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_F8B8BS_BH_SAB_UserArgs_shortname0_gfx950 DebugStreamK: 0 DepthU: 128 DirectToLds: true @@ -231,83 +231,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -397,7 +320,7 @@ ClusterLocalRead: 0 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_F8B8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_F8B8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950 DebugStreamK: 0 DepthU: 128 DirectToLds: true @@ -531,83 +454,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -697,7 +543,7 @@ ClusterLocalRead: 0 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_F8B8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_F8B8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950 DebugStreamK: 0 DepthU: 128 DirectToLds: true @@ -831,83 +677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -997,7 +766,7 @@ ClusterLocalRead: 0 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_F8B8BS_BH_SAB_NTD_UserArgs_shortname1_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_F8B8BS_BH_SAB_NTD_UserArgs_shortname1_gfx950 DebugStreamK: 0 DepthU: 128 DirectToLds: true @@ -1131,83 +900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1297,7 +989,7 @@ ClusterLocalRead: 0 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_F8B8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_F8B8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950 DebugStreamK: 0 DepthU: 128 DirectToLds: true @@ -1431,83 +1123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1581,17 +1196,17 @@ - [2, 3, 0, 1] - - - [28672, 49152, 1, 8192] - [3, 0] - - - [8192, 28672, 1, 49152] + - - [8192, 28672, 1, 49152] - [3, 0] - - - [8192, 49152, 1, 28672] + - - [8192, 49152, 1, 28672] - [3, 0] - - - [8192, 49152, 1, 8192] + - - [8192, 49152, 1, 8192] - [3, 0] - - - [28672, 8192, 1, 49152] + - - [28672, 8192, 1, 49152] - [3, 0] - - - [8192, 8192, 1, 49152] + - - [8192, 8192, 1, 49152] - [3, 0] - - - [1024, 49152, 1, 8192] + - - [1024, 49152, 1, 8192] - [4, 0] - - [57344, 32768, 1, 8192] - [0, 0] @@ -1653,7 +1268,6 @@ - [3, 0] - - [4096, 4096, 1, 32768] - [2, 0] - - - [14336, 65536, 1, 4096] - [1, 0] - - [14336, 4096, 1, 65536] @@ -1668,7 +1282,6 @@ - [1, 0] - - [4096, 4096, 1, 65536] - [2, 0] - - - [14336, 4096, 1, 49152] - [0, 0] - - [14336, 49152, 1, 4096] diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_F8BS_BH_SAB_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_F8BS_BH_SAB_UserArgs.yaml index d48dec69343..3fa4dd81613 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_F8BS_BH_SAB_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_F8BS_BH_SAB_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -97,7 +97,7 @@ ClusterLocalRead: 0 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_F8BS_BH_SAB_UserArgs_shortname0_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_F8BS_BH_SAB_UserArgs_shortname0_gfx950 DebugStreamK: 0 DepthU: 128 DirectToLds: true @@ -231,83 +231,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -397,7 +320,7 @@ ClusterLocalRead: 0 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_F8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_F8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950 DebugStreamK: 0 DepthU: 128 DirectToLds: true @@ -531,83 +454,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -697,7 +543,7 @@ ClusterLocalRead: 0 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_F8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_F8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950 DebugStreamK: 0 DepthU: 128 DirectToLds: true @@ -831,83 +677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -997,7 +766,7 @@ ClusterLocalRead: 0 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_F8BS_BH_SAB_NTD_UserArgs_shortname1_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_F8BS_BH_SAB_NTD_UserArgs_shortname1_gfx950 DebugStreamK: 0 DepthU: 128 DirectToLds: true @@ -1131,83 +900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1297,7 +989,7 @@ ClusterLocalRead: 0 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_F8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_F8BS_BH_SAB_NTD_UserArgs_shortname0_gfx950 DebugStreamK: 0 DepthU: 128 DirectToLds: true @@ -1431,83 +1123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1581,17 +1196,17 @@ - [2, 3, 0, 1] - - - [28672, 49152, 1, 8192] - [3, 0] - - - [8192, 28672, 1, 49152] + - - [8192, 28672, 1, 49152] - [3, 0] - - - [8192, 49152, 1, 28672] + - - [8192, 49152, 1, 28672] - [1, 0] - - - [8192, 49152, 1, 8192] + - - [8192, 49152, 1, 8192] - [2, 0] - - - [28672, 8192, 1, 49152] + - - [28672, 8192, 1, 49152] - [3, 0] - - - [8192, 8192, 1, 49152] + - - [8192, 8192, 1, 49152] - [3, 0] - - - [1024, 49152, 1, 8192] + - - [1024, 49152, 1, 8192] - [4, 0] - - [57344, 32768, 1, 8192] - [3, 0] @@ -1653,7 +1268,6 @@ - [2, 0] - - [4096, 4096, 1, 32768] - [2, 0] - - - [4096, 65536, 1, 4096] - [2, 0] - - [2048, 65536, 1, 4096] @@ -1662,7 +1276,6 @@ - [1, 0] - - [4096, 65536, 1, 14336] - [2, 0] - - - [14336, 4096, 1, 49152] - [0, 0] - - [14336, 49152, 1, 4096] diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_HHS_BH_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_HHS_BH_UserArgs.yaml index 370cdf5edaa..f73ac8bb83c 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_HHS_BH_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/Equality/gfx950_Cijk_Alik_Bljk_HHS_BH_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -97,7 +97,7 @@ ClusterLocalRead: 1 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_HHS_BH_MT256x256x64_MI16x16x1_UserArgs_shortname0_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_HHS_BH_MT256x256x64_MI16x16x1_UserArgs_shortname0_gfx950 DebugStreamK: 0 DepthU: 64 DirectToLds: true @@ -233,83 +233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -399,7 +322,7 @@ ClusterLocalRead: 1 CodeObjectVersion: '4' ConvertAfterDS: false - CustomKernelName: 'Custom_Cijk_Alik_Bljk_HHS_BH_MT256x256x64_MI16x16x1_UserArgs_shortname0_gfx950' + CustomKernelName: Custom_Cijk_Alik_Bljk_HHS_BH_MT256x256x64_MI16x16x1_UserArgs_shortname0_gfx950 DebugStreamK: 0 DepthU: 64 DirectToLds: true @@ -535,83 +458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: false - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BBS_BH_Bias_BiasSrcB_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BBS_BH_Bias_BiasSrcB_SAV_UserArgs.yaml index f7edb81970d..5ea1d5f606d 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BBS_BH_Bias_BiasSrcB_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BBS_BH_Bias_BiasSrcB_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -70,7 +70,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 UseScaleCD: false - - 1LDSBuffer: 0 @@ -205,78 +205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -462,78 +390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -719,78 +575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -976,78 +760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1235,79 +947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1496,79 +1135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BBS_BH_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BBS_BH_Bias_HAS_SAV_UserArgs.yaml index a69386f629f..1c187976541 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BBS_BH_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BBS_BH_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -222,81 +222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -506,81 +431,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -790,81 +640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1074,81 +849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1358,81 +1058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1642,81 +1267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1926,81 +1476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2210,81 +1685,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2494,81 +1894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2778,81 +2103,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3062,81 +2312,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3346,81 +2521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3630,81 +2730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3914,81 +2939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4198,81 +3148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4482,81 +3357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4766,81 +3566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5050,81 +3775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5334,81 +3984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5618,81 +4193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5902,81 +4402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6186,81 +4611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6470,81 +4820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6754,81 +5029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7038,81 +5238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7322,81 +5447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7606,81 +5656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7890,81 +5865,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8174,81 +6074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8458,81 +6283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8742,81 +6492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9026,81 +6701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9310,81 +6910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9594,81 +7119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9878,81 +7328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10162,81 +7537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10446,81 +7746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10730,81 +7955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11014,81 +8164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11298,81 +8373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11582,81 +8582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11866,81 +8791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12150,81 +9000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12434,81 +9209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12718,81 +9418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13002,81 +9627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13286,81 +9836,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13570,81 +10045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13854,81 +10254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14138,81 +10463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14422,81 +10672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14706,81 +10881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14990,81 +11090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15274,81 +11299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15558,81 +11508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15842,81 +11717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16126,81 +11926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16410,81 +12135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16694,81 +12344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16978,81 +12553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17262,81 +12762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17546,81 +12971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17830,81 +13180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18114,81 +13389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18398,81 +13598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18682,81 +13807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18966,81 +14016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19250,81 +14225,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19534,81 +14434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19818,81 +14643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20102,81 +14852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20386,81 +15061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20670,81 +15270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20954,81 +15479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21238,81 +15688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21522,81 +15897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21806,81 +16106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22090,81 +16315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22374,81 +16524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22658,81 +16733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22942,81 +16942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23226,81 +17151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23510,81 +17360,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23794,81 +17569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24078,81 +17778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24362,81 +17987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24646,81 +18196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24930,81 +18405,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25214,81 +18614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25498,81 +18823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25782,81 +19032,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26066,81 +19241,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26350,81 +19450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26634,81 +19659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26918,81 +19868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27202,81 +20077,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27486,81 +20286,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27770,81 +20495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28054,81 +20704,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28338,81 +20913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28622,81 +21122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28906,81 +21331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29190,81 +21540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29474,81 +21749,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29758,81 +21958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30042,81 +22167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30326,81 +22376,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30610,81 +22585,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30902,83 +22802,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31203,83 +23026,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31504,83 +23250,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31805,83 +23474,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32106,83 +23698,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32407,83 +23922,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32708,83 +24146,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33009,83 +24370,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33310,83 +24594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33611,83 +24818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33912,83 +25042,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34213,83 +25266,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34514,83 +25490,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34815,83 +25714,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35116,83 +25938,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35417,83 +26162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35718,83 +26386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36019,83 +26610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36320,83 +26834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36621,83 +27058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36922,83 +27282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37223,83 +27506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37524,83 +27730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37825,83 +27954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38126,83 +28178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38428,83 +28403,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38729,83 +28627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39030,83 +28851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39332,83 +29076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39633,83 +29300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39934,83 +29524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40235,83 +29748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40536,83 +29972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40837,83 +30196,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41138,83 +30420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41439,83 +30644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41741,83 +30869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42042,83 +31093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42344,83 +31318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42645,83 +31542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42946,83 +31766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43247,83 +31990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43548,83 +32214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43849,83 +32438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44151,83 +32663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44452,83 +32887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44753,83 +33111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45054,83 +33335,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45356,83 +33560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45657,83 +33784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45959,83 +34009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46261,83 +34234,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46563,83 +34459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46864,83 +34683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47165,83 +34907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47466,83 +35131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47767,83 +35355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48068,83 +35579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48369,83 +35803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48671,83 +36028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48973,83 +36253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49274,83 +36477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49575,83 +36701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49877,83 +36926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50178,83 +37150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50479,83 +37374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50781,83 +37599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51082,83 +37823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51383,83 +38047,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51684,83 +38271,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51985,83 +38495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52286,83 +38719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52587,83 +38943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52888,83 +39167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53189,83 +39391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53490,83 +39615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53792,83 +39840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54093,83 +40064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54394,83 +40288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54695,83 +40512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54996,83 +40736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55297,83 +40960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55598,83 +41184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55899,83 +41408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56200,83 +41632,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56501,83 +41856,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56803,83 +42081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57105,83 +42306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57406,83 +42530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57707,83 +42754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58009,83 +42979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58310,83 +43203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58611,83 +43427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58912,83 +43651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59214,83 +43876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59515,83 +44100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59816,83 +44324,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60117,83 +44548,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60418,83 +44772,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60719,83 +44996,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61020,83 +45220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61321,83 +45444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61623,83 +45669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61924,83 +45893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62226,83 +46118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62527,83 +46342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62829,83 +46567,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63130,83 +46791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63431,83 +47015,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63733,83 +47240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64034,83 +47464,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64335,83 +47688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64637,83 +47913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64938,83 +48137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65239,83 +48361,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65540,83 +48585,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65842,83 +48810,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66144,83 +49035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66446,83 +49260,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66747,83 +49484,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67048,83 +49708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67349,83 +49932,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67650,83 +50156,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67951,83 +50380,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68253,83 +50605,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68554,83 +50829,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68855,83 +51053,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69156,83 +51277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69457,83 +51501,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69759,83 +51726,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70061,83 +51951,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70362,83 +52175,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70663,83 +52399,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70964,83 +52623,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71265,83 +52847,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71566,83 +53071,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71867,83 +53295,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72168,83 +53519,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72469,83 +53743,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72770,83 +53967,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73071,83 +54191,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73372,83 +54415,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73673,83 +54639,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73975,83 +54864,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74276,83 +55088,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74577,83 +55312,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74878,83 +55536,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75179,83 +55760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75480,83 +55984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75781,83 +56208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76082,83 +56432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76383,83 +56656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76684,83 +56880,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76985,83 +57104,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77286,83 +57328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77587,83 +57552,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77888,83 +57776,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78189,83 +58000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78490,83 +58224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78792,83 +58449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79093,83 +58673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79394,83 +58897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79695,83 +59121,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79996,83 +59345,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80297,83 +59569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80598,83 +59793,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80899,83 +60017,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81200,83 +60241,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81501,83 +60465,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81802,83 +60689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82103,83 +60913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82404,83 +61137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82705,83 +61361,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83006,83 +61585,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83307,83 +61809,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83608,83 +62033,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83910,83 +62258,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84211,83 +62482,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84513,83 +62707,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84815,83 +62932,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85116,83 +63156,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85418,83 +63381,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85719,83 +63605,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86020,83 +63829,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86321,83 +64053,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86622,83 +64277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86923,83 +64501,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87224,83 +64725,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87525,83 +64949,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87826,83 +65173,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88128,83 +65398,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88429,83 +65622,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88731,83 +65847,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89033,83 +66072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89334,83 +66296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89635,83 +66520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89936,83 +66744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90238,83 +66969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90539,83 +67193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90840,83 +67417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91141,83 +67641,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91442,83 +67865,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91744,83 +68090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92045,83 +68314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92346,83 +68538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92647,83 +68762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92948,83 +68986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93250,83 +69211,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93551,83 +69435,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93852,83 +69659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94153,83 +69883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94454,83 +70107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94756,83 +70332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95057,83 +70556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95359,83 +70781,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95661,83 +71006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95962,83 +71230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96263,83 +71454,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96564,83 +71678,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96866,83 +71903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97167,83 +72127,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97469,83 +72352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97770,83 +72576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98071,83 +72800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98372,83 +73024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98673,83 +73248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98974,83 +73472,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99275,83 +73696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99577,83 +73921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99878,83 +74145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100180,83 +74370,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100481,83 +74594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100782,83 +74818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101083,83 +75042,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101384,83 +75266,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101685,83 +75490,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101986,83 +75714,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102287,83 +75938,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102589,83 +76163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102890,83 +76387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103191,83 +76611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103492,83 +76835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103793,83 +77059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104094,83 +77283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104395,83 +77507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104696,83 +77731,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104997,83 +77955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105298,83 +78179,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105600,83 +78404,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105901,83 +78628,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106202,83 +78852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106503,83 +79076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106804,83 +79300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107105,83 +79524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107406,83 +79748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107707,83 +79972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108009,83 +80197,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108310,83 +80421,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108612,83 +80646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108914,83 +80871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109215,83 +81095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109516,83 +81319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109817,83 +81543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110118,83 +81767,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110420,83 +81992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110721,83 +82216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111022,83 +82440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111324,83 +82665,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111626,83 +82890,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111927,83 +83114,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112228,83 +83338,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112529,83 +83562,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112830,83 +83786,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113132,83 +84011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113434,83 +84236,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113736,83 +84461,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114038,83 +84686,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114340,83 +84911,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114642,83 +85136,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114943,83 +85360,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115244,83 +85584,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115545,83 +85808,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115846,83 +86032,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116147,83 +86256,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116448,83 +86480,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116750,83 +86705,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117052,83 +86930,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117353,83 +87154,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117654,83 +87378,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117955,83 +87602,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118257,83 +87827,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118558,83 +88051,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118859,83 +88275,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119160,83 +88499,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119461,83 +88723,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119762,83 +88947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120064,83 +89172,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120365,83 +89396,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120666,83 +89620,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120968,83 +89845,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121270,83 +90070,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121571,83 +90294,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121873,83 +90519,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122174,83 +90743,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122476,83 +90968,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122778,83 +91193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123079,83 +91417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123380,83 +91641,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123681,83 +91865,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123982,83 +92089,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124284,83 +92314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124585,83 +92538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124886,83 +92762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125188,83 +92987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125489,83 +93211,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125790,83 +93435,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126091,83 +93659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126393,83 +93884,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126695,83 +94109,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126996,83 +94333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127297,83 +94557,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127598,83 +94781,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127899,83 +95005,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128200,83 +95229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128501,83 +95453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128802,83 +95677,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129103,83 +95901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129404,83 +96125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129705,83 +96349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130006,83 +96573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130307,83 +96797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130608,83 +97021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130909,83 +97245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131210,83 +97469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131511,83 +97693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131812,83 +97917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132114,83 +98142,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132415,83 +98366,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132716,83 +98590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133017,83 +98814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133318,83 +99038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133619,83 +99262,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133920,83 +99486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134221,83 +99710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134522,83 +99934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134823,83 +100158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135124,83 +100382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135426,83 +100607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135727,83 +100831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136028,83 +101055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136330,83 +101280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136632,83 +101505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136934,83 +101730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137235,83 +101954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137536,83 +102178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137837,83 +102402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138139,83 +102627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138440,83 +102851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138741,83 +103075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139042,83 +103299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139344,83 +103524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139645,83 +103748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139946,83 +103972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140247,83 +104196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140548,83 +104420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140850,83 +104645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141151,83 +104869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141453,83 +105094,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141755,83 +105319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142056,83 +105543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142358,83 +105768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142659,83 +105992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142960,83 +106216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143261,83 +106440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143562,83 +106664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143863,83 +106888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144164,83 +107112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144466,83 +107337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144768,83 +107562,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145069,83 +107786,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145370,83 +108010,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145671,83 +108234,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145972,83 +108458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146273,83 +108682,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146574,83 +108906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146875,83 +109130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147177,83 +109355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147478,83 +109579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147779,83 +109803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148080,83 +110027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148382,83 +110252,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148684,83 +110477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148985,83 +110701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149286,83 +110925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149587,83 +111149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149888,83 +111373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150190,83 +111598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150491,83 +111822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150792,83 +112046,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151093,83 +112270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151394,83 +112494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151695,83 +112718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151996,83 +112942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152298,83 +113167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152600,83 +113392,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152901,83 +113616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153202,83 +113840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153503,83 +114064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153804,83 +114288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154105,83 +114512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154406,83 +114736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154708,83 +114961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155009,83 +115185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155311,83 +115410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155612,83 +115634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155913,83 +115858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156214,83 +116082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156516,83 +116307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156817,83 +116531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157118,83 +116755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157419,83 +116979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157720,83 +117203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158022,83 +117428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158323,83 +117652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158625,83 +117877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158926,83 +118101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159227,83 +118325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159528,83 +118549,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159829,83 +118773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160130,83 +118997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160431,83 +119221,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160732,83 +119445,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161033,83 +119669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161335,83 +119894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161636,83 +120118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161937,83 +120342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162238,83 +120566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162539,83 +120790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162840,83 +121014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163141,83 +121238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163443,83 +121463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163744,83 +121687,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164045,83 +121911,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164346,83 +122135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164647,83 +122359,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164948,83 +122583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165249,83 +122807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165550,83 +123031,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165851,83 +123255,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166152,83 +123479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166454,83 +123704,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166755,83 +123928,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167056,83 +124152,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167357,83 +124376,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167659,83 +124601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167960,83 +124825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168261,83 +125049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168562,83 +125273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168863,83 +125497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169164,83 +125721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169465,83 +125945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169766,83 +126169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170067,83 +126393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170368,83 +126617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170669,83 +126841,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170970,83 +127065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171271,83 +127289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171572,83 +127513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171873,83 +127737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172174,83 +127961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172475,83 +128185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172776,83 +128409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173078,83 +128634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173380,83 +128859,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173682,83 +129084,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173983,83 +129308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174284,83 +129532,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174585,83 +129756,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174886,83 +129980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175187,83 +130204,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175488,83 +130428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175790,83 +130653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176091,83 +130877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176392,83 +131101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176693,83 +131325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176994,83 +131549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177295,83 +131773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177597,83 +131998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177898,83 +132222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178199,83 +132446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178500,83 +132670,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178802,83 +132895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179103,83 +133119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179404,83 +133343,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179705,83 +133567,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180006,83 +133791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180307,83 +134015,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180608,83 +134239,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180909,83 +134463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181210,83 +134687,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181511,83 +134911,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181812,83 +135135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182113,83 +135359,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182414,83 +135583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182715,83 +135807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183016,83 +136031,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183317,83 +136255,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183618,83 +136479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183919,83 +136703,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184220,83 +136927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184521,83 +137151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184822,83 +137375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185123,83 +137599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185424,83 +137823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185725,83 +138047,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186027,83 +138272,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186329,83 +138497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186630,83 +138721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186931,83 +138945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187232,83 +139169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187533,83 +139393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187835,83 +139618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188136,83 +139842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188437,83 +140066,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188738,83 +140290,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189039,83 +140514,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189340,83 +140738,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189642,83 +140963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189943,83 +141187,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190244,83 +141411,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190545,83 +141635,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190847,83 +141860,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191148,83 +142084,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191449,83 +142308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191750,83 +142532,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192052,83 +142757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192354,83 +142982,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192655,83 +143206,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192956,83 +143430,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193257,83 +143654,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193559,83 +143879,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193860,83 +144103,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194162,83 +144328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194464,83 +144553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194765,83 +144777,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195066,83 +145001,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195367,83 +145225,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195668,83 +145449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195969,83 +145673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196271,83 +145898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196572,83 +146122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196874,83 +146347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197176,83 +146572,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197478,83 +146797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197779,83 +147021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198080,83 +147245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198381,83 +147469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198682,83 +147693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198984,83 +147918,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199286,83 +148143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199587,83 +148367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199889,83 +148592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200190,83 +148816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200491,83 +149040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200792,83 +149264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201093,83 +149488,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201394,83 +149712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201695,83 +149936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201996,83 +150160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202297,83 +150384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202598,83 +150608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202899,83 +150832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203200,83 +151056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203502,83 +151281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203803,83 +151505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204104,83 +151729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204405,83 +151953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204706,83 +152177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205008,83 +152402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205310,83 +152627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205611,83 +152851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205913,83 +153076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206214,83 +153300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206515,83 +153524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206816,83 +153748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207117,83 +153972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207418,83 +154196,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207719,83 +154420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208020,83 +154644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208321,83 +154868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208622,83 +155092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208923,83 +155316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209224,83 +155540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209526,83 +155765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209828,83 +155990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210129,83 +156214,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210431,83 +156439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210732,83 +156663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211033,83 +156887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211334,83 +157111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211635,83 +157335,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211936,83 +157559,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212237,83 +157783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212538,83 +158007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212839,83 +158231,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213140,83 +158455,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213441,83 +158679,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213742,83 +158903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214043,83 +159127,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214344,83 +159351,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214646,83 +159576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214947,83 +159800,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215248,83 +160024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215549,83 +160248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215850,83 +160472,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216151,83 +160696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216452,83 +160920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216753,83 +161144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217054,83 +161368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217356,83 +161593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217657,83 +161817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217958,83 +162041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218259,83 +162265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218560,83 +162489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218862,83 +162714,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219163,83 +162938,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219464,83 +163162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219765,83 +163386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220067,83 +163611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220368,83 +163835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220670,83 +164060,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220971,83 +164284,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221272,83 +164508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221573,83 +164732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221874,83 +164956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222175,83 +165180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222476,83 +165404,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222777,83 +165628,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223078,83 +165852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223379,83 +166076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223680,83 +166300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223982,83 +166525,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224284,83 +166750,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224585,83 +166974,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224886,83 +167198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225187,83 +167422,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225488,83 +167646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225790,83 +167871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226091,83 +168095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226392,83 +168319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226693,83 +168543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226994,83 +168767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227295,83 +168991,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227596,83 +169215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227898,83 +169440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228200,83 +169665,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228501,83 +169889,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228802,83 +170113,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229104,83 +170338,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229405,83 +170562,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229707,83 +170787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230008,83 +171011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230309,83 +171235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230610,83 +171459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230912,83 +171684,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231214,83 +171909,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231515,83 +172133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231816,83 +172357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232117,83 +172581,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232419,83 +172806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232720,83 +173030,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233022,83 +173255,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233323,83 +173479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233625,83 +173704,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233926,83 +173928,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234227,83 +174152,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234528,83 +174376,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234829,83 +174600,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235131,83 +174825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235432,83 +175049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235733,83 +175273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236034,83 +175497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236336,83 +175722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236637,83 +175946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236938,83 +176170,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237240,83 +176395,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237542,83 +176620,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237843,83 +176844,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238144,83 +177068,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238445,83 +177292,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238746,83 +177516,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239048,83 +177741,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239349,83 +177965,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239650,83 +178189,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239952,83 +178414,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240254,83 +178639,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240555,83 +178863,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240856,83 +179087,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241158,83 +179312,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241460,83 +179537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241761,83 +179761,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242062,83 +179985,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242363,83 +180209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242665,83 +180434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242967,83 +180659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243268,83 +180883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243569,83 +181107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243871,83 +181332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244173,83 +181557,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244474,83 +181781,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244775,83 +182005,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245076,83 +182229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245377,83 +182453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245679,83 +182678,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245981,83 +182903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246282,83 +183127,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246584,83 +183352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246885,83 +183576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247187,83 +183801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247488,83 +184025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247790,83 +184250,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248092,83 +184475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248393,83 +184699,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248694,83 +184923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248995,83 +185147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249296,83 +185371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249597,83 +185595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249899,83 +185820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250201,83 +186045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250503,83 +186270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250805,83 +186495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251106,83 +186719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251407,83 +186943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251708,83 +187167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252009,83 +187391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252310,83 +187615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252612,83 +187840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252913,83 +188064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253214,83 +188288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253515,83 +188512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253816,83 +188736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254117,83 +188960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BSS_BH_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BSS_BH_Bias_HAS_SAV_UserArgs.yaml index 34686fa161b..54169695ede 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BSS_BH_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_BSS_BH_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -834,83 +680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1135,83 +904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1436,83 +1128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1737,83 +1352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2039,83 +1577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2340,83 +1801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2641,83 +2025,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2942,83 +2249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3243,83 +2473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3544,83 +2697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3845,83 +2921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4146,83 +3145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4447,83 +3369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4748,83 +3593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5049,83 +3817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5350,83 +4041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5651,83 +4265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5952,83 +4489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6253,83 +4713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6554,83 +4937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6856,83 +5162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7157,83 +5386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7458,83 +5610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7759,83 +5834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8060,83 +6058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8361,83 +6282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8662,83 +6506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8963,83 +6730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9264,83 +6954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9565,83 +7178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9866,83 +7402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10168,83 +7627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10469,83 +7851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10770,83 +8075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11071,83 +8299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11372,83 +8523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11673,83 +8747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11974,83 +8971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12275,83 +9195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12576,83 +9419,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12877,83 +9643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13178,83 +9867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13479,83 +10091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13780,83 +10315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14081,83 +10539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14382,83 +10763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14683,83 +10987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14984,83 +11211,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15285,83 +11435,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15586,83 +11659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15887,83 +11883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16188,83 +12107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16489,83 +12331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16790,83 +12555,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17091,83 +12779,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17393,83 +13004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17694,83 +13228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17995,83 +13452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18296,83 +13676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18597,83 +13900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18899,83 +14125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19201,83 +14350,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19503,83 +14575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19804,83 +14799,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20106,83 +15024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20408,83 +15249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20710,83 +15474,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21011,83 +15698,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21313,83 +15923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21614,83 +16147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21915,83 +16371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22216,83 +16595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22518,83 +16820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22820,83 +17045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23121,83 +17269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23422,83 +17493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23724,83 +17718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24025,83 +17942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24326,83 +18166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24628,83 +18391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24929,83 +18615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25230,83 +18839,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25531,83 +19063,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25832,83 +19287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26133,83 +19511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26434,83 +19735,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26735,83 +19959,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27036,83 +20183,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27337,83 +20407,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27639,83 +20632,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27941,83 +20857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28242,83 +21081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28543,83 +21305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28845,83 +21530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29146,83 +21754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29448,83 +21979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29749,83 +22203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30051,83 +22428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30352,83 +22652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30654,83 +22877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30956,83 +23102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31257,83 +23326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31559,83 +23551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31861,83 +23776,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32162,83 +24000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32463,83 +24224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32764,83 +24448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33065,83 +24672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33366,83 +24896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33667,83 +25120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33968,83 +25344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34270,83 +25569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34572,83 +25794,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34874,83 +26019,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35176,83 +26244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35477,83 +26468,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35778,83 +26692,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36079,83 +26916,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36380,83 +27140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36681,83 +27364,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36982,83 +27588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37283,83 +27812,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37584,83 +28036,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37885,83 +28260,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38187,83 +28485,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38488,83 +28709,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38789,83 +28933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39090,83 +29157,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39391,83 +29381,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39692,83 +29605,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39994,83 +29830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40295,83 +30054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40596,83 +30278,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40897,83 +30502,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41198,83 +30726,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41499,83 +30950,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41800,83 +31174,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42101,83 +31398,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42402,83 +31622,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42703,83 +31846,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43004,83 +32070,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43305,83 +32294,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43606,83 +32518,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43907,83 +32742,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44208,83 +32966,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44509,83 +33190,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44810,83 +33414,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45111,83 +33638,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45412,83 +33862,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45713,83 +34086,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46014,83 +34310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46315,83 +34534,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46616,83 +34758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46918,83 +34983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47220,83 +35208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47521,83 +35432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47823,83 +35657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48124,83 +35881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48426,83 +36106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48728,83 +36331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49029,83 +36555,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49330,83 +36779,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49631,83 +37003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49932,83 +37227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50233,83 +37451,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50534,83 +37675,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50835,83 +37899,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51136,83 +38123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51437,83 +38347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51738,83 +38571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52039,83 +38795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52340,83 +39019,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52641,83 +39243,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52942,83 +39467,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53243,83 +39691,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53544,83 +39915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53845,83 +40139,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54146,83 +40363,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54448,83 +40588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54750,83 +40813,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55051,83 +41037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55352,83 +41261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55654,83 +41486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55955,83 +41710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56256,83 +41934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56557,83 +42158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56858,83 +42382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57159,83 +42606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57460,83 +42830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57762,83 +43055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58063,83 +43279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58364,83 +43503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58666,83 +43728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58967,83 +43952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59268,83 +44176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59569,83 +44400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59870,83 +44624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60171,83 +44848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60473,83 +45073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60775,83 +45298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61076,83 +45522,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61377,83 +45746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61679,83 +45971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61980,83 +46195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62281,83 +46419,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62582,83 +46643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62883,83 +46867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63184,83 +47091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63485,83 +47315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63787,83 +47540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64088,83 +47764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64390,83 +47989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64692,83 +48214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64993,83 +48438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65294,83 +48662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65596,83 +48887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65898,83 +49112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66200,83 +49337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66502,83 +49562,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66804,83 +49787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67105,83 +50011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67407,83 +50236,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67709,83 +50461,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68010,83 +50685,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68312,83 +50910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68613,83 +51134,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68914,83 +51358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69215,83 +51582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69516,83 +51806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69817,83 +52030,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70118,83 +52254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70419,83 +52478,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70720,83 +52702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71022,83 +52927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71324,83 +53152,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71625,83 +53376,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71926,83 +53600,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72227,83 +53824,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72529,83 +54049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72830,83 +54273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73132,83 +54498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73434,83 +54723,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73735,83 +54947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74036,83 +55171,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74337,83 +55395,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74638,83 +55619,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74940,83 +55844,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75241,83 +56068,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75542,83 +56292,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75843,83 +56516,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76144,83 +56740,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76445,83 +56964,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76746,83 +57188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77047,83 +57412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77348,83 +57636,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77649,83 +57860,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77951,83 +58085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78252,83 +58309,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78554,83 +58534,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78855,83 +58758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79156,83 +58982,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79457,83 +59206,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79759,83 +59431,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80061,83 +59656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80363,83 +59881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80665,83 +60106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80966,83 +60330,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81267,83 +60554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81568,83 +60778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81869,83 +61002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82170,83 +61226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82472,83 +61451,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82774,83 +61676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83075,83 +61900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83377,83 +62125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83678,83 +62349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83979,83 +62573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84281,83 +62798,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84582,83 +63022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84883,83 +63246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85184,83 +63470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85486,83 +63695,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85788,83 +63920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86089,83 +64144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86390,83 +64368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86692,83 +64593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86994,83 +64818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87296,83 +65043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87598,83 +65268,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87900,83 +65493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88201,83 +65717,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88502,83 +65941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88803,83 +66165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89104,83 +66389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89406,83 +66614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89707,83 +66838,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90008,83 +67062,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90309,83 +67286,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90610,83 +67510,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90911,83 +67734,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91212,83 +67958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91513,83 +68182,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91814,83 +68406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92115,83 +68630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92417,83 +68855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92719,83 +69080,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93021,83 +69305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93322,83 +69529,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93623,83 +69753,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93925,83 +69978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94227,83 +70203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94528,83 +70427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94830,83 +70652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95131,83 +70876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95432,83 +71100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95733,83 +71324,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96034,83 +71548,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96336,83 +71773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96637,83 +71997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96939,83 +72222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97240,83 +72446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97541,83 +72670,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97842,83 +72894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98143,83 +73118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98444,83 +73342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98745,83 +73566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99046,83 +73790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99347,83 +74014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99649,83 +74239,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99950,83 +74463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100251,83 +74687,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100553,83 +74912,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100854,83 +75136,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101155,83 +75360,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101456,83 +75584,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101757,83 +75808,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102058,83 +76032,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102360,83 +76257,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102661,83 +76481,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102963,83 +76706,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103265,83 +76931,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103566,83 +77155,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103867,83 +77379,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104168,83 +77603,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104469,83 +77827,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104770,83 +78051,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105071,83 +78275,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105372,83 +78499,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105673,83 +78723,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105974,83 +78947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106275,83 +79171,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106576,83 +79395,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106877,83 +79619,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107179,83 +79844,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107481,83 +80069,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107782,83 +80293,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108083,83 +80517,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108384,83 +80741,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108685,83 +80965,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108986,83 +81189,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109288,83 +81414,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109590,83 +81639,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109891,83 +81863,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110192,83 +82087,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110493,83 +82311,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110794,83 +82535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111095,83 +82759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111397,83 +82984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111698,83 +83208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111999,83 +83432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112300,83 +83656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112601,83 +83880,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112902,83 +84104,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113203,83 +84328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113505,83 +84553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113807,83 +84778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114108,83 +85002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114409,83 +85226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114710,83 +85450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115011,83 +85674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115312,83 +85898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115614,83 +86123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115916,83 +86348,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116217,83 +86572,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116518,83 +86796,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116819,83 +87020,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117120,83 +87244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117422,83 +87469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117724,83 +87694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118026,83 +87919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118327,83 +88143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118628,83 +88367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118929,83 +88591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119230,83 +88815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119531,83 +89039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119832,83 +89263,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120133,83 +89487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120434,83 +89711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120735,83 +89935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121036,83 +90159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121337,83 +90383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121639,83 +90608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121941,83 +90833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122242,83 +91057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122543,83 +91281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122844,83 +91505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123146,83 +91730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123447,83 +91954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123748,83 +92178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124049,83 +92402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124350,83 +92626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124651,83 +92850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124952,83 +93074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125253,83 +93298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125555,83 +93523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125856,83 +93747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126158,83 +93972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126459,83 +94196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126760,83 +94420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127061,83 +94644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127362,83 +94868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127664,83 +95093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127965,83 +95317,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128266,83 +95541,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128567,83 +95765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128868,83 +95989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129170,83 +96214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129471,83 +96438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129772,83 +96662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130073,83 +96886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130374,83 +97110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130676,83 +97335,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130977,83 +97559,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131278,83 +97783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131579,83 +98007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131880,83 +98231,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132181,83 +98455,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132482,83 +98679,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132783,83 +98903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133084,83 +99127,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133385,83 +99351,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133686,83 +99575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133988,83 +99800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134289,83 +100024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134590,83 +100248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134891,83 +100472,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135193,83 +100697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135494,83 +100921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135795,83 +101145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136096,83 +101369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136398,83 +101594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136699,83 +101818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137001,83 +102043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137302,83 +102267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137603,83 +102491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137904,83 +102715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138206,83 +102940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138508,83 +103165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138810,83 +103390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139111,83 +103614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139412,83 +103838,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139713,83 +104062,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140014,83 +104286,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140315,83 +104510,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140616,83 +104734,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140917,83 +104958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141219,83 +105183,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141520,83 +105407,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141821,83 +105631,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142122,83 +105855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142423,83 +106079,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142724,83 +106303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143025,83 +106527,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143326,83 +106751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143627,83 +106975,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143928,83 +107199,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144229,83 +107423,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144531,83 +107648,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144832,83 +107872,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145133,83 +108096,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145435,83 +108321,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145737,83 +108546,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146039,83 +108771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146340,83 +108995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146641,83 +109219,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146942,83 +109443,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147243,83 +109667,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147544,83 +109891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147845,83 +110115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148147,83 +110340,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148449,83 +110565,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148751,83 +110790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149053,83 +111015,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149355,83 +111240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149656,83 +111464,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149958,83 +111689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150259,83 +111913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150561,83 +112138,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150863,83 +112363,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151164,83 +112587,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151466,83 +112812,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151767,83 +113036,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152068,83 +113260,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152370,83 +113485,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152671,83 +113709,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152973,83 +113934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153274,83 +114158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153575,83 +114382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153877,83 +114607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154178,83 +114831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154479,83 +115055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154780,83 +115279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155081,83 +115503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155382,83 +115727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155684,83 +115952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155985,83 +116176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156287,83 +116401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156589,83 +116626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156890,83 +116850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157192,83 +117075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157493,83 +117299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157794,83 +117523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158095,83 +117747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158396,83 +117971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158697,83 +118195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158998,83 +118419,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159299,83 +118643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159601,83 +118868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159902,83 +119092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160203,83 +119316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160505,83 +119541,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160806,83 +119765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161107,83 +119989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161408,83 +120213,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161710,83 +120438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162011,83 +120662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162312,83 +120886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162613,83 +121110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162914,83 +121334,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163215,83 +121558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163516,83 +121782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163818,83 +122007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164119,83 +122231,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164420,83 +122455,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164721,83 +122679,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165022,83 +122903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165323,83 +123127,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165624,83 +123351,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165926,83 +123576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166227,83 +123800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166528,83 +124024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166829,83 +124248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167130,83 +124472,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167431,83 +124696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167732,83 +124920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168033,83 +125144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168335,83 +125369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168636,83 +125593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168937,83 +125817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169238,83 +126041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169539,83 +126265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169841,83 +126490,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170143,83 +126715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170445,83 +126940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170746,83 +127164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171048,83 +127389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171349,83 +127613,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171650,83 +127837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171952,83 +128062,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172253,83 +128286,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172555,83 +128511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172857,83 +128736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173158,83 +128960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173460,83 +129185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173761,83 +129409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174062,83 +129633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174363,83 +129857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174664,83 +130081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174965,83 +130305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175266,83 +130529,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175567,83 +130753,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175869,83 +130978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176171,83 +131203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176473,83 +131428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176775,83 +131653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177077,83 +131878,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177378,83 +132102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177679,83 +132326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177980,83 +132550,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178282,83 +132775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178584,83 +133000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178885,83 +133224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179187,83 +133449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179488,83 +133673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179789,83 +133897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180091,83 +134122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180392,83 +134346,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180693,83 +134570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180995,83 +134795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181297,83 +135020,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181599,83 +135245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181901,83 +135470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182202,83 +135694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182503,83 +135918,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182804,83 +136142,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183105,83 +136366,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183407,83 +136591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183708,83 +136815,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184009,83 +137039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184311,83 +137264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184612,83 +137488,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184913,83 +137712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185214,83 +137936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185515,83 +138160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185817,83 +138385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186119,83 +138610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186421,83 +138835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186722,83 +139059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187023,83 +139283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187325,83 +139508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187626,83 +139732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187928,83 +139957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188229,83 +140181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188530,83 +140405,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188831,83 +140629,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189132,83 +140853,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189433,83 +141077,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189734,83 +141301,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190035,83 +141525,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190336,83 +141749,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190637,83 +141973,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190939,83 +142198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191240,83 +142422,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191541,83 +142646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191842,83 +142870,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192144,83 +143095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192445,83 +143319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192746,83 +143543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193048,83 +143768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193349,83 +143992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193650,83 +144216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193951,83 +144440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194252,83 +144664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194553,83 +144888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194854,83 +145112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195155,83 +145336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195456,83 +145560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195758,83 +145785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196059,83 +146009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196360,83 +146233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196662,83 +146458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196963,83 +146682,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197264,83 +146906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197566,83 +147131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197867,83 +147355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198168,83 +147579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198469,83 +147803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198770,83 +148027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199071,83 +148251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199372,83 +148475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199673,83 +148699,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199974,83 +148923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200275,83 +149147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200577,83 +149372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200878,83 +149596,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201180,83 +149821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201481,83 +150045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201783,83 +150270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202085,83 +150495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202386,83 +150719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202687,83 +150943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202988,83 +151167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203289,83 +151391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203591,83 +151616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203892,83 +151840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204193,83 +152064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204495,83 +152289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204797,83 +152514,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205099,83 +152739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205400,83 +152963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205702,83 +153188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206003,83 +153412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206304,83 +153636,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206605,83 +153860,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206906,83 +154084,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207207,83 +154308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207509,83 +154533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207810,83 +154757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208112,83 +154982,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208414,83 +155207,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208715,83 +155431,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209017,83 +155656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209318,83 +155880,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209619,83 +156104,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209920,83 +156328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210221,83 +156552,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210522,83 +156776,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210823,83 +157000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211124,83 +157224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211426,83 +157449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211727,83 +157673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212029,83 +157898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212330,83 +158122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212631,83 +158346,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212932,83 +158570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213233,83 +158794,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213534,83 +159018,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213835,83 +159242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214136,83 +159466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214437,83 +159690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214738,83 +159914,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215040,83 +160139,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215342,83 +160364,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215643,83 +160588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215944,83 +160812,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216246,83 +161037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216547,83 +161261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216849,83 +161486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217150,83 +161710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217451,83 +161934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217752,83 +162158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218053,83 +162382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218354,83 +162606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218656,83 +162831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218958,83 +163056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219259,83 +163280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219561,83 +163505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219863,83 +163730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220164,83 +163954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220465,83 +164178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220767,83 +164403,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221068,83 +164627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221370,83 +164852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221672,83 +165077,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221974,83 +165302,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222275,83 +165526,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222576,83 +165750,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222877,83 +165974,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223178,83 +166198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223479,83 +166422,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223781,83 +166647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224083,83 +166872,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224384,83 +167096,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224685,83 +167320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224986,83 +167544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225288,83 +167769,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225590,83 +167994,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225891,83 +168218,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226192,83 +168442,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226493,83 +168666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226794,83 +168890,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227095,83 +169114,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227396,83 +169338,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227698,83 +169563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228000,83 +169788,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228301,83 +170012,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228602,83 +170236,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228904,83 +170461,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229205,83 +170685,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229506,83 +170909,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229807,83 +171133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230108,83 +171357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230409,83 +171581,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230710,83 +171805,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231012,83 +172030,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231313,83 +172254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231614,83 +172478,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231915,83 +172702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232216,83 +172926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232517,83 +173150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232818,83 +173374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233120,83 +173599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233421,83 +173823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233722,83 +174047,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234024,83 +174272,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234325,83 +174496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234627,83 +174721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234928,83 +174945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235229,83 +175169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235530,83 +175393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235831,83 +175617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236132,83 +175841,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236433,83 +176065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236734,83 +176289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237035,83 +176513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237336,83 +176737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237637,83 +176961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237938,83 +177185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238239,83 +177409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238540,83 +177633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238841,83 +177857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239142,83 +178081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239443,83 +178305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239745,83 +178530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240047,83 +178755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240349,83 +178980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240650,83 +179204,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240951,83 +179428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241252,83 +179652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241553,83 +179876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241854,83 +180100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242155,83 +180324,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242456,83 +180548,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242758,83 +180773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243059,83 +180997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243360,83 +181221,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243661,83 +181445,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243963,83 +181670,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244264,83 +181894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244566,83 +182119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244867,83 +182343,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245168,83 +182567,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245469,83 +182791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245770,83 +183015,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246071,83 +183239,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246372,83 +183463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246674,83 +183688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246975,83 +183912,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247276,83 +184136,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247578,83 +184361,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247879,83 +184585,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248180,83 +184809,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248481,83 +185033,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248782,83 +185257,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249083,83 +185481,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249384,83 +185705,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249685,83 +185929,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249986,83 +186153,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250287,83 +186377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250588,83 +186601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250889,83 +186825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251190,83 +187049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251491,83 +187273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251792,83 +187497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252093,83 +187721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252395,83 +187946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252696,83 +188170,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252998,83 +188395,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253299,83 +188619,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253601,83 +188844,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253902,83 +189068,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254204,83 +189293,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254505,83 +189517,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254806,83 +189741,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255107,83 +189965,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255409,83 +190190,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255710,83 +190414,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256011,83 +190638,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256312,83 +190862,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256613,83 +191086,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256915,83 +191311,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257216,83 +191535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257517,83 +191759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257818,83 +191983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258120,83 +192208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258421,83 +192432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258722,83 +192656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -259024,83 +192881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -259325,83 +193105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -259626,83 +193329,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -259927,83 +193553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -260229,83 +193778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -260530,83 +194002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -260831,83 +194226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -261132,83 +194450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -261433,83 +194674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -261734,83 +194898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -262035,83 +195122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -262336,83 +195346,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -262638,83 +195571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -262940,83 +195796,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -263242,83 +196021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -263543,83 +196245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_DB_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_DB_UserArgs.yaml index e96dab878d8..7b0d7bd284b 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_DB_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_DB_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -67,7 +67,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 0 - - 1LDSBuffer: 0 ActivationAlt: false @@ -194,74 +194,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 1 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 1 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - DataType: 1 - DestDataType: 1 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 0 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -440,74 +372,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 1 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 1 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - DataType: 1 - DestDataType: 1 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 0 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8B8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8B8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml index 15d966fe873..1fc537b1f7b 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8B8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8B8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -834,83 +680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1135,83 +904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1436,83 +1128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1737,83 +1352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2038,83 +1576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2339,83 +1800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2640,83 +2024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2942,83 +2249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3243,83 +2473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3544,83 +2697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3845,83 +2921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4146,83 +3145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4447,83 +3369,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4748,83 +3593,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5049,83 +3817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5350,83 +4041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5651,83 +4265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5952,83 +4489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6253,83 +4713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6554,83 +4937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6855,83 +5161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7156,83 +5385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7457,83 +5609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7758,83 +5833,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8059,83 +6057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8360,83 +6281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8661,83 +6505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8962,83 +6729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9263,83 +6953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9564,83 +7177,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9865,83 +7401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10166,83 +7625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10467,83 +7849,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10768,83 +8073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11069,83 +8297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11370,83 +8521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11671,83 +8745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11972,83 +8969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12273,83 +9193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12574,83 +9417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12876,83 +9642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13178,83 +9867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13479,83 +10091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13780,83 +10315,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14081,83 +10539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14382,83 +10763,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14683,83 +10987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14985,83 +11212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15286,83 +11436,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15587,83 +11660,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15888,83 +11884,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16189,83 +12108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16490,83 +12332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16791,83 +12556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17092,83 +12780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17393,83 +13004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17694,83 +13228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17995,83 +13452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18296,83 +13676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18597,83 +13900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18898,83 +14124,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19200,83 +14349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19501,83 +14573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19803,83 +14798,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20104,83 +15022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20405,83 +15246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20706,83 +15470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21008,83 +15695,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21309,83 +15919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21610,83 +16143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21911,83 +16367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22212,83 +16591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22513,83 +16815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22814,83 +17039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23115,83 +17263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23416,83 +17487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23717,83 +17711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24018,83 +17935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24319,83 +18159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24620,83 +18383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24921,83 +18607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25222,83 +18831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25523,83 +19055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25824,83 +19279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26125,83 +19503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26426,83 +19727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26727,83 +19951,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27029,83 +20176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27331,83 +20401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27633,83 +20626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27935,83 +20851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28236,83 +21075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28537,83 +21299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28838,83 +21523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29139,83 +21747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29440,83 +21971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29742,83 +22196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30043,83 +22420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30344,83 +22644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30645,83 +22868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30946,83 +23092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31247,83 +23316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31548,83 +23540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31849,83 +23764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32151,83 +23989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32452,83 +24213,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32753,83 +24437,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33054,83 +24661,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33355,83 +24885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33656,83 +25109,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33957,83 +25333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34258,83 +25557,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34559,83 +25781,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34860,83 +26005,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35161,83 +26229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35462,83 +26453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35763,83 +26677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36064,83 +26901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36365,83 +27125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36666,83 +27349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36967,83 +27573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37268,83 +27797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37569,83 +28021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37870,83 +28245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38171,83 +28469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38472,83 +28693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38773,83 +28917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39074,83 +29141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39376,83 +29366,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39677,83 +29590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39978,83 +29814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40279,83 +30038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40581,83 +30263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40882,83 +30487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41184,83 +30712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41485,83 +30936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41786,83 +31160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42087,83 +31384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42388,83 +31608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42689,83 +31832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42990,83 +32056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43291,83 +32280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43593,83 +32505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43894,83 +32729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44195,83 +32953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44497,83 +33178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44798,83 +33402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45099,83 +33626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45400,83 +33850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45702,83 +34075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46004,83 +34300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46305,83 +34524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46606,83 +34748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46907,83 +34972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47208,83 +35196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47510,83 +35421,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47812,83 +35646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48114,83 +35871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48415,83 +36095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48717,83 +36320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49018,83 +36544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49319,83 +36768,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49620,83 +36992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49921,83 +37216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50222,83 +37440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50523,83 +37664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50824,83 +37888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51125,83 +38112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51426,83 +38336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51727,83 +38560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52028,83 +38784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52329,83 +39008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52630,83 +39232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52931,83 +39456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53232,83 +39680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53533,83 +39904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53834,83 +40128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54135,83 +40352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54436,83 +40576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54737,83 +40800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55038,83 +41024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55340,83 +41249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55642,83 +41474,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55943,83 +41698,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56245,83 +41923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56546,83 +42147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56847,83 +42371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57148,83 +42595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57449,83 +42819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57750,83 +43043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58051,83 +43267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58352,83 +43491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58653,83 +43715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58954,83 +43939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59255,83 +44163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59556,83 +44387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59857,83 +44611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60158,83 +44835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60460,83 +45060,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60762,83 +45285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61063,83 +45509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61364,83 +45733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61665,83 +45957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61966,83 +46181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62268,83 +46406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62569,83 +46630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62871,83 +46855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63172,83 +47079,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63473,83 +47303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63774,83 +47527,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64075,83 +47751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64376,83 +47975,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64678,83 +48200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64979,83 +48424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65280,83 +48648,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65581,83 +48872,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65882,83 +49096,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66183,83 +49320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66484,83 +49544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66785,83 +49768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67086,83 +49992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67387,83 +50216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67689,83 +50441,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67991,83 +50666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68292,83 +50890,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68594,83 +51115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68895,83 +51339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69196,83 +51563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69497,83 +51787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69798,83 +52011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70099,83 +52235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70400,83 +52459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70701,83 +52683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71002,83 +52907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71303,83 +53131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71604,83 +53355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71906,83 +53580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72208,83 +53805,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72509,83 +54029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72811,83 +54254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73113,83 +54479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73415,83 +54704,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73716,83 +54928,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74018,83 +55153,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74319,83 +55377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74620,83 +55601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74921,83 +55825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75222,83 +56049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75523,83 +56273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75824,83 +56497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76125,83 +56721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76426,83 +56945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76727,83 +57169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77028,83 +57393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77329,83 +57617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77630,83 +57841,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77931,83 +58065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78233,83 +58290,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78534,83 +58514,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78836,83 +58739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79137,83 +58963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79439,83 +59188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79740,83 +59412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80042,83 +59637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80344,83 +59862,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80646,83 +60087,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80947,83 +60311,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81248,83 +60535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81550,83 +60760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81851,83 +60984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82152,83 +61208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82454,83 +61433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82755,83 +61657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83056,83 +61881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83357,83 +62105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83658,83 +62329,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83959,83 +62553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84260,83 +62777,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84561,83 +63001,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84863,83 +63226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85164,83 +63450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85465,83 +63674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85766,83 +63898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86068,83 +64123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86369,83 +64347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86671,83 +64572,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86973,83 +64797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87274,83 +65021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87575,83 +65245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87876,83 +65469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88177,83 +65693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88478,83 +65917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88779,83 +66141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89080,83 +66365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89382,83 +66590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89683,83 +66814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89984,83 +67038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90286,83 +67263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90587,83 +67487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90888,83 +67711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91189,83 +67935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91491,83 +68160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91792,83 +68384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92093,83 +68608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92394,83 +68832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92695,83 +69056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92996,83 +69280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93297,83 +69504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93598,83 +69728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93899,83 +69952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94200,83 +70176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94502,83 +70401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94803,83 +70625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95104,83 +70849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95405,83 +71073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95707,83 +71298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96009,83 +71523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96311,83 +71748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96612,83 +71972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96913,83 +72196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97214,83 +72420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97516,83 +72645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97817,83 +72869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98118,83 +73093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98419,83 +73317,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98721,83 +73542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99022,83 +73766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99323,83 +73990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99624,83 +74214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99925,83 +74438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100226,83 +74662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100527,83 +74886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100828,83 +75110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101129,83 +75334,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101430,83 +75558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101731,83 +75782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102033,83 +76007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102335,83 +76232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102636,83 +76456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102937,83 +76680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103238,83 +76904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103540,83 +77129,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103842,83 +77354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104143,83 +77578,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104445,83 +77803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104747,83 +78028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105049,83 +78253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105350,83 +78477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105652,83 +78702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105953,83 +78926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106254,83 +79150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106555,83 +79374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106856,83 +79598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107157,83 +79822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107458,83 +80046,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107759,83 +80270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108061,83 +80495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108363,83 +80720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108665,83 +80945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108966,83 +81169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109268,83 +81394,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109569,83 +81618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109870,83 +81842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110172,83 +82067,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110473,83 +82291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110775,83 +82516,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111076,83 +82740,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111377,83 +82964,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111678,83 +83188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111979,83 +83412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112281,83 +83637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112582,83 +83861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112884,83 +84086,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113185,83 +84310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113487,83 +84535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113789,83 +84760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114090,83 +84984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114391,83 +85208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114693,83 +85433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114994,83 +85657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115295,83 +85881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115596,83 +86105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115897,83 +86329,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116198,83 +86553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116500,83 +86778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116802,83 +87003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117103,83 +87227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117405,83 +87452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117706,83 +87676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118007,83 +87900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118309,83 +88125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118610,83 +88349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118911,83 +88573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119212,83 +88797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119514,83 +89022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119815,83 +89246,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120116,83 +89470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120417,83 +89694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120719,83 +89919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121020,83 +90143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121322,83 +90368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121623,83 +90592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121924,83 +90816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122225,83 +91040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122526,83 +91264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122828,83 +91489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123129,83 +91713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123430,83 +91937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123732,83 +92162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124033,83 +92386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124334,83 +92610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124635,83 +92834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124936,83 +93058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125237,83 +93282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125538,83 +93506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125840,83 +93731,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126141,83 +93955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126442,83 +94179,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126743,83 +94403,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127044,83 +94627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127345,83 +94851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127647,83 +95076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127948,83 +95300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128249,83 +95524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128551,83 +95749,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128853,83 +95974,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129154,83 +96198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129456,83 +96423,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129757,83 +96647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130058,83 +96871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130359,83 +97095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130660,83 +97319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130961,83 +97543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131262,83 +97767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131563,83 +97991,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131864,83 +98215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132165,83 +98439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132467,83 +98664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132769,83 +98889,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133071,83 +99114,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133372,83 +99338,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133673,83 +99562,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133974,83 +99786,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134276,83 +100011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134577,83 +100235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134878,83 +100459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135179,83 +100683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135480,83 +100907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135781,83 +101131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136082,83 +101355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136383,83 +101579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136684,83 +101803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136985,83 +102027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137286,83 +102251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137587,83 +102475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137888,83 +102699,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138189,83 +102923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138490,83 +103147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138792,83 +103372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139094,83 +103597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139395,83 +103821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139696,83 +104045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139997,83 +104269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140298,83 +104493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140600,83 +104718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140901,83 +104942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141203,83 +105167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141504,83 +105391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141806,83 +105616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142107,83 +105840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142408,83 +106064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142709,83 +106288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143010,83 +106512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143311,83 +106736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143613,83 +106961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143914,83 +107185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144215,83 +107409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144516,83 +107633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144818,83 +107858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145119,83 +108082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145420,83 +108306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145722,83 +108531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146024,83 +108756,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146325,83 +108980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146626,83 +109204,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146927,83 +109428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147228,83 +109652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147530,83 +109877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147831,83 +110101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148133,83 +110326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148435,83 +110551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148736,83 +110775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149037,83 +110999,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149339,83 +111224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149640,83 +111448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149941,83 +111672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150242,83 +111896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150543,83 +112120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150844,83 +112344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151145,83 +112568,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151446,83 +112792,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151747,83 +113016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152048,83 +113240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152350,83 +113465,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152651,83 +113689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152952,83 +113913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153253,83 +114137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153555,83 +114362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153856,83 +114586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154157,83 +114810,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154459,83 +115035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154760,83 +115259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155062,83 +115484,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155363,83 +115708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155665,83 +115933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155966,83 +116157,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156267,83 +116381,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156569,83 +116606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156870,83 +116830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157171,83 +117054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157473,83 +117279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157774,83 +117503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158075,83 +117727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158376,83 +117951,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158678,83 +118176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158979,83 +118400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159280,83 +118624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159581,83 +118848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159882,83 +119072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160183,83 +119296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160484,83 +119520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160785,83 +119744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161087,83 +119969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161388,83 +120193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161689,83 +120417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161991,83 +120642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162292,83 +120866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162593,83 +121090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162894,83 +121314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163195,83 +121538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163496,83 +121762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163797,83 +121986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164099,83 +122211,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164400,83 +122435,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164701,83 +122659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165002,83 +122883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165304,83 +123108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165605,83 +123332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165907,83 +123557,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166209,83 +123782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166510,83 +124006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166811,83 +124230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167113,83 +124455,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167415,83 +124680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167716,83 +124904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168017,83 +125128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168319,83 +125353,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168620,83 +125577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168921,83 +125801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169222,83 +126025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169523,83 +126249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169824,83 +126473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170125,83 +126697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170426,83 +126921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170727,83 +127145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171029,83 +127370,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171331,83 +127595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171632,83 +127819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171933,83 +128043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172234,83 +128267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172535,83 +128491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172836,83 +128715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173137,83 +128939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173438,83 +129163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173739,83 +129387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174041,83 +129612,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174343,83 +129837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174644,83 +130061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174945,83 +130285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175246,83 +130509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175547,83 +130733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175849,83 +130958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176150,83 +131182,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176451,83 +131406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176753,83 +131631,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177054,83 +131855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177355,83 +132079,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177656,83 +132303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177958,83 +132528,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178260,83 +132753,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178561,83 +132977,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178862,83 +133201,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179163,83 +133425,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179464,83 +133649,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179766,83 +133874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180068,83 +134099,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180369,83 +134323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180670,83 +134547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180971,83 +134771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181273,83 +134996,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181574,83 +135220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181875,83 +135444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182176,83 +135668,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182478,83 +135893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182780,83 +136118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183081,83 +136342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183382,83 +136566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183683,83 +136790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183984,83 +137014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml index e4871ba98bf..49673115612 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -834,83 +680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1135,83 +904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1436,83 +1128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1737,83 +1352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2038,83 +1576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2339,83 +1800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2640,83 +2024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2942,83 +2249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3243,83 +2473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3544,83 +2697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3845,83 +2921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4146,83 +3145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4447,83 +3369,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4748,83 +3593,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5049,83 +3817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5350,83 +4041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5651,83 +4265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5952,83 +4489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6253,83 +4713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6554,83 +4937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6855,83 +5161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7156,83 +5385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7457,83 +5609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7758,83 +5833,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8059,83 +6057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8360,83 +6281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8661,83 +6505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8962,83 +6729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9263,83 +6953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9564,83 +7177,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9865,83 +7401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10166,83 +7625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10467,83 +7849,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10768,83 +8073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11069,83 +8297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11370,83 +8521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11671,83 +8745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11972,83 +8969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12273,83 +9193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12574,83 +9417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12876,83 +9642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13178,83 +9867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13479,83 +10091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13780,83 +10315,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14081,83 +10539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14382,83 +10763,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14683,83 +10987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14985,83 +11212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15286,83 +11436,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15587,83 +11660,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15888,83 +11884,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16189,83 +12108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16490,83 +12332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16791,83 +12556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17092,83 +12780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17393,83 +13004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17694,83 +13228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17995,83 +13452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18296,83 +13676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18597,83 +13900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18898,83 +14124,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19200,83 +14349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19501,83 +14573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19803,83 +14798,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20104,83 +15022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20405,83 +15246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20706,83 +15470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21008,83 +15695,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21309,83 +15919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21610,83 +16143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21911,83 +16367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22212,83 +16591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22513,83 +16815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22814,83 +17039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23115,83 +17263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23416,83 +17487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23717,83 +17711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24018,83 +17935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24319,83 +18159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24620,83 +18383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24921,83 +18607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25222,83 +18831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25523,83 +19055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25824,83 +19279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26125,83 +19503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26426,83 +19727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26727,83 +19951,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27029,83 +20176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27331,83 +20401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27633,83 +20626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27935,83 +20851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28236,83 +21075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28537,83 +21299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28838,83 +21523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29139,83 +21747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29440,83 +21971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29742,83 +22196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30043,83 +22420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30344,83 +22644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30645,83 +22868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30946,83 +23092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31247,83 +23316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31548,83 +23540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31849,83 +23764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32151,83 +23989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32452,83 +24213,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32753,83 +24437,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33054,83 +24661,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33355,83 +24885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33656,83 +25109,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33957,83 +25333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34258,83 +25557,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34559,83 +25781,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34860,83 +26005,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35161,83 +26229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35462,83 +26453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35763,83 +26677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36064,83 +26901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36365,83 +27125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36666,83 +27349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36967,83 +27573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37268,83 +27797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37569,83 +28021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37870,83 +28245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38171,83 +28469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38472,83 +28693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38773,83 +28917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39074,83 +29141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39376,83 +29366,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39677,83 +29590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39978,83 +29814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40279,83 +30038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40581,83 +30263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40882,83 +30487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41184,83 +30712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41485,83 +30936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41786,83 +31160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42087,83 +31384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42388,83 +31608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42689,83 +31832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42990,83 +32056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43291,83 +32280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43593,83 +32505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43894,83 +32729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44195,83 +32953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44497,83 +33178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44798,83 +33402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45099,83 +33626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45400,83 +33850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45702,83 +34075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46004,83 +34300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46305,83 +34524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46606,83 +34748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46907,83 +34972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47208,83 +35196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47510,83 +35421,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47812,83 +35646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48114,83 +35871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48415,83 +36095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48717,83 +36320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49018,83 +36544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49319,83 +36768,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49620,83 +36992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49921,83 +37216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50222,83 +37440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50523,83 +37664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50824,83 +37888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51125,83 +38112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51426,83 +38336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51727,83 +38560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52028,83 +38784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52329,83 +39008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52630,83 +39232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52931,83 +39456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53232,83 +39680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53533,83 +39904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53834,83 +40128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54135,83 +40352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54436,83 +40576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54737,83 +40800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55038,83 +41024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55340,83 +41249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55642,83 +41474,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55943,83 +41698,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56245,83 +41923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56546,83 +42147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56847,83 +42371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57148,83 +42595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57449,83 +42819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57750,83 +43043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58051,83 +43267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58352,83 +43491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58653,83 +43715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58954,83 +43939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59255,83 +44163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59556,83 +44387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59857,83 +44611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60158,83 +44835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60460,83 +45060,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60762,83 +45285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61063,83 +45509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61364,83 +45733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61665,83 +45957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61966,83 +46181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62268,83 +46406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62569,83 +46630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62871,83 +46855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63172,83 +47079,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63473,83 +47303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63774,83 +47527,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64075,83 +47751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64376,83 +47975,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64678,83 +48200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64979,83 +48424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65280,83 +48648,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65581,83 +48872,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65882,83 +49096,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66183,83 +49320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66484,83 +49544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66785,83 +49768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67086,83 +49992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67387,83 +50216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67689,83 +50441,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67991,83 +50666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68292,83 +50890,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68594,83 +51115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68895,83 +51339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69196,83 +51563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69497,83 +51787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69798,83 +52011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70099,83 +52235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70400,83 +52459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70701,83 +52683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71002,83 +52907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71303,83 +53131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71604,83 +53355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71906,83 +53580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72208,83 +53805,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72509,83 +54029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72811,83 +54254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73113,83 +54479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73415,83 +54704,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73716,83 +54928,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74018,83 +55153,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74319,83 +55377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74620,83 +55601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74921,83 +55825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75222,83 +56049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75523,83 +56273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75824,83 +56497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76125,83 +56721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76426,83 +56945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76727,83 +57169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77028,83 +57393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77329,83 +57617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77630,83 +57841,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77931,83 +58065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78233,83 +58290,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78534,83 +58514,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78836,83 +58739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79137,83 +58963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79439,83 +59188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79740,83 +59412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80042,83 +59637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80344,83 +59862,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80646,83 +60087,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80947,83 +60311,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81248,83 +60535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81550,83 +60760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81851,83 +60984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82152,83 +61208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82454,83 +61433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82755,83 +61657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83056,83 +61881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83357,83 +62105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83658,83 +62329,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83959,83 +62553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84260,83 +62777,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84561,83 +63001,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84863,83 +63226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85164,83 +63450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85465,83 +63674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85766,83 +63898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86068,83 +64123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86369,83 +64347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86671,83 +64572,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86973,83 +64797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87274,83 +65021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87575,83 +65245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87876,83 +65469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88177,83 +65693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88478,83 +65917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88779,83 +66141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89080,83 +66365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89382,83 +66590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89683,83 +66814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89984,83 +67038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90286,83 +67263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90587,83 +67487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90888,83 +67711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91189,83 +67935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91491,83 +68160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91792,83 +68384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92093,83 +68608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92394,83 +68832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92695,83 +69056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92996,83 +69280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93297,83 +69504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93598,83 +69728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93899,83 +69952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94200,83 +70176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94502,83 +70401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94803,83 +70625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95104,83 +70849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95405,83 +71073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95707,83 +71298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96009,83 +71523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96311,83 +71748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96612,83 +71972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96913,83 +72196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97214,83 +72420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97516,83 +72645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97817,83 +72869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98118,83 +73093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98419,83 +73317,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98721,83 +73542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99022,83 +73766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99323,83 +73990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99624,83 +74214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99925,83 +74438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100226,83 +74662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100527,83 +74886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100828,83 +75110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101129,83 +75334,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101430,83 +75558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101731,83 +75782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102033,83 +76007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102335,83 +76232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102636,83 +76456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102937,83 +76680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103238,83 +76904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103540,83 +77129,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103842,83 +77354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104143,83 +77578,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104445,83 +77803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104747,83 +78028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105049,83 +78253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105350,83 +78477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105652,83 +78702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105953,83 +78926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106254,83 +79150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106555,83 +79374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106856,83 +79598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107157,83 +79822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107458,83 +80046,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107759,83 +80270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108061,83 +80495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108363,83 +80720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108665,83 +80945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108966,83 +81169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109268,83 +81394,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109569,83 +81618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109870,83 +81842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110172,83 +82067,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110473,83 +82291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110775,83 +82516,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111076,83 +82740,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111377,83 +82964,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111678,83 +83188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111979,83 +83412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112281,83 +83637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112582,83 +83861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112884,83 +84086,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113185,83 +84310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113487,83 +84535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113789,83 +84760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114090,83 +84984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114391,83 +85208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114693,83 +85433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114994,83 +85657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115295,83 +85881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115596,83 +86105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115897,83 +86329,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116198,83 +86553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116500,83 +86778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116802,83 +87003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117103,83 +87227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117405,83 +87452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117706,83 +87676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118007,83 +87900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118309,83 +88125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118610,83 +88349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118911,83 +88573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119212,83 +88797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119514,83 +89022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119815,83 +89246,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120116,83 +89470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120417,83 +89694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120719,83 +89919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121020,83 +90143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121322,83 +90368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121623,83 +90592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121924,83 +90816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122225,83 +91040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122526,83 +91264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122828,83 +91489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123129,83 +91713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123430,83 +91937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123732,83 +92162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124033,83 +92386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124334,83 +92610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124635,83 +92834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124936,83 +93058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125237,83 +93282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125538,83 +93506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125840,83 +93731,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126141,83 +93955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126442,83 +94179,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126743,83 +94403,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127044,83 +94627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127345,83 +94851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127647,83 +95076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127948,83 +95300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128249,83 +95524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128551,83 +95749,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128853,83 +95974,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129154,83 +96198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129456,83 +96423,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129757,83 +96647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130058,83 +96871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130359,83 +97095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130660,83 +97319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130961,83 +97543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131262,83 +97767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131563,83 +97991,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131864,83 +98215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132165,83 +98439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132467,83 +98664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132769,83 +98889,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133071,83 +99114,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133372,83 +99338,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133673,83 +99562,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133974,83 +99786,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134276,83 +100011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134577,83 +100235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134878,83 +100459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135179,83 +100683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135480,83 +100907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135781,83 +101131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136082,83 +101355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136383,83 +101579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136684,83 +101803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136985,83 +102027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137286,83 +102251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137587,83 +102475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137888,83 +102699,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138189,83 +102923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138490,83 +103147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138792,83 +103372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139094,83 +103597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139395,83 +103821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139696,83 +104045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139997,83 +104269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140298,83 +104493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140600,83 +104718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140901,83 +104942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141203,83 +105167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141504,83 +105391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141806,83 +105616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142107,83 +105840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142408,83 +106064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142709,83 +106288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143010,83 +106512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143311,83 +106736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143613,83 +106961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143914,83 +107185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144215,83 +107409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144516,83 +107633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144818,83 +107858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145119,83 +108082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145420,83 +108306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145722,83 +108531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146024,83 +108756,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146325,83 +108980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146626,83 +109204,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146927,83 +109428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147228,83 +109652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147530,83 +109877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147831,83 +110101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148133,83 +110326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148435,83 +110551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148736,83 +110775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149037,83 +110999,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149339,83 +111224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149640,83 +111448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149941,83 +111672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150242,83 +111896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150543,83 +112120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150844,83 +112344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151145,83 +112568,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151446,83 +112792,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151747,83 +113016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152048,83 +113240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152350,83 +113465,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152651,83 +113689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152952,83 +113913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153253,83 +114137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153555,83 +114362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153856,83 +114586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154157,83 +114810,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154459,83 +115035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154760,83 +115259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155062,83 +115484,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155363,83 +115708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155665,83 +115933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155966,83 +116157,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156267,83 +116381,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156569,83 +116606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156870,83 +116830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157171,83 +117054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157473,83 +117279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157774,83 +117503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158075,83 +117727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158376,83 +117951,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158678,83 +118176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158979,83 +118400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159280,83 +118624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159581,83 +118848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159882,83 +119072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160183,83 +119296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160484,83 +119520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160785,83 +119744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161087,83 +119969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161388,83 +120193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161689,83 +120417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161991,83 +120642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162292,83 +120866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162593,83 +121090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162894,83 +121314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163195,83 +121538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163496,83 +121762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163797,83 +121986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164099,83 +122211,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164400,83 +122435,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164701,83 +122659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165002,83 +122883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165304,83 +123108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165605,83 +123332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165907,83 +123557,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166209,83 +123782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166510,83 +124006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166811,83 +124230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167113,83 +124455,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167415,83 +124680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167716,83 +124904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168017,83 +125128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168319,83 +125353,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168620,83 +125577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168921,83 +125801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169222,83 +126025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169523,83 +126249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169824,83 +126473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170125,83 +126697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170426,83 +126921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170727,83 +127145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171029,83 +127370,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171331,83 +127595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171632,83 +127819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171933,83 +128043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172234,83 +128267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172535,83 +128491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172836,83 +128715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173137,83 +128939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173438,83 +129163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173739,83 +129387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174041,83 +129612,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174343,83 +129837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174644,83 +130061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174945,83 +130285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175246,83 +130509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175547,83 +130733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175849,83 +130958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176150,83 +131182,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176451,83 +131406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176753,83 +131631,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177054,83 +131855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177355,83 +132079,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177656,83 +132303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177958,83 +132528,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178260,83 +132753,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178561,83 +132977,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178862,83 +133201,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179163,83 +133425,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179464,83 +133649,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179766,83 +133874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180068,83 +134099,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180369,83 +134323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180670,83 +134547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180971,83 +134771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181273,83 +134996,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181574,83 +135220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181875,83 +135444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182176,83 +135668,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182478,83 +135893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182780,83 +136118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183081,83 +136342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183382,83 +136566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183683,83 +136790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183984,83 +137014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8HS_BH_BiasSH_HAS_SAB_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8HS_BH_BiasSH_HAS_SAB_SAV_UserArgs.yaml index c5ae0a3d269..989adf1f047 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8HS_BH_BiasSH_HAS_SAB_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_F8HS_BH_BiasSH_HAS_SAB_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -834,83 +680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1135,83 +904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1436,83 +1128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1737,83 +1352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2038,83 +1576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2339,83 +1800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2640,83 +2024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2942,83 +2249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3243,83 +2473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3544,83 +2697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3845,83 +2921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4146,83 +3145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4447,83 +3369,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4748,83 +3593,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5049,83 +3817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5350,83 +4041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5651,83 +4265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5952,83 +4489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6253,83 +4713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6554,83 +4937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6855,83 +5161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7156,83 +5385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7457,83 +5609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7758,83 +5833,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8059,83 +6057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8360,83 +6281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8661,83 +6505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8962,83 +6729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9263,83 +6953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9564,83 +7177,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9865,83 +7401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10166,83 +7625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10467,83 +7849,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10768,83 +8073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11069,83 +8297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11370,83 +8521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11671,83 +8745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11972,83 +8969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12273,83 +9193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12574,83 +9417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12876,83 +9642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13178,83 +9867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13479,83 +10091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13780,83 +10315,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14081,83 +10539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14382,83 +10763,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14683,83 +10987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14985,83 +11212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15286,83 +11436,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15587,83 +11660,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15888,83 +11884,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16189,83 +12108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16490,83 +12332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16791,83 +12556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17092,83 +12780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17393,83 +13004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17694,83 +13228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17995,83 +13452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18296,83 +13676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18597,83 +13900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18898,83 +14124,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19200,83 +14349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19501,83 +14573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19803,83 +14798,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20104,83 +15022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20405,83 +15246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20706,83 +15470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21008,83 +15695,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21309,83 +15919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21610,83 +16143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21911,83 +16367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22212,83 +16591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22513,83 +16815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22814,83 +17039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23115,83 +17263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23416,83 +17487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23717,83 +17711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24018,83 +17935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24319,83 +18159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24620,83 +18383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24921,83 +18607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25222,83 +18831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25523,83 +19055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25824,83 +19279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26125,83 +19503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26426,83 +19727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26727,83 +19951,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27029,83 +20176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27331,83 +20401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27633,83 +20626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27935,83 +20851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28236,83 +21075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28537,83 +21299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28838,83 +21523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29139,83 +21747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29440,83 +21971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29742,83 +22196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30043,83 +22420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30344,83 +22644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30645,83 +22868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30946,83 +23092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31247,83 +23316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31548,83 +23540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31849,83 +23764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32151,83 +23989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32452,83 +24213,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32753,83 +24437,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33054,83 +24661,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33355,83 +24885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33656,83 +25109,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33957,83 +25333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34258,83 +25557,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34559,83 +25781,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34860,83 +26005,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35161,83 +26229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35462,83 +26453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35763,83 +26677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36064,83 +26901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36365,83 +27125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36666,83 +27349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36967,83 +27573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37268,83 +27797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37569,83 +28021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37870,83 +28245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38171,83 +28469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38472,83 +28693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38773,83 +28917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39074,83 +29141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39376,83 +29366,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39677,83 +29590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39978,83 +29814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40279,83 +30038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40581,83 +30263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40882,83 +30487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41184,83 +30712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41485,83 +30936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41786,83 +31160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42087,83 +31384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42388,83 +31608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42689,83 +31832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42990,83 +32056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43291,83 +32280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43593,83 +32505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43894,83 +32729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44195,83 +32953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44497,83 +33178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44798,83 +33402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45099,83 +33626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45400,83 +33850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45702,83 +34075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46004,83 +34300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46305,83 +34524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46606,83 +34748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46907,83 +34972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47208,83 +35196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47510,83 +35421,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47812,83 +35646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48114,83 +35871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48415,83 +36095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48717,83 +36320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49018,83 +36544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49319,83 +36768,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49620,83 +36992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49921,83 +37216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50222,83 +37440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50523,83 +37664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50824,83 +37888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51125,83 +38112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51426,83 +38336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51727,83 +38560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52028,83 +38784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52329,83 +39008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52630,83 +39232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52931,83 +39456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53232,83 +39680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53533,83 +39904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53834,83 +40128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54135,83 +40352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54436,83 +40576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54737,83 +40800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55038,83 +41024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55340,83 +41249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55642,83 +41474,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55943,83 +41698,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56245,83 +41923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56546,83 +42147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56847,83 +42371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57148,83 +42595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57449,83 +42819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57750,83 +43043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58051,83 +43267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58352,83 +43491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58653,83 +43715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58954,83 +43939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59255,83 +44163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59556,83 +44387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59857,83 +44611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60158,83 +44835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60460,83 +45060,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60762,83 +45285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61063,83 +45509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61364,83 +45733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61665,83 +45957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61966,83 +46181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62268,83 +46406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62569,83 +46630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62871,83 +46855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63172,83 +47079,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63473,83 +47303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63774,83 +47527,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64075,83 +47751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64376,83 +47975,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64678,83 +48200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64979,83 +48424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65280,83 +48648,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65581,83 +48872,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65882,83 +49096,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66183,83 +49320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66484,83 +49544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66785,83 +49768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67086,83 +49992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67387,83 +50216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67689,83 +50441,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67991,83 +50666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68292,83 +50890,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68594,83 +51115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68895,83 +51339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69196,83 +51563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69497,83 +51787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69798,83 +52011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70099,83 +52235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70400,83 +52459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70701,83 +52683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71002,83 +52907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71303,83 +53131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71604,83 +53355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71906,83 +53580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72208,83 +53805,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72509,83 +54029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72811,83 +54254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73113,83 +54479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73415,83 +54704,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73716,83 +54928,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74018,83 +55153,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74319,83 +55377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74620,83 +55601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74921,83 +55825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75222,83 +56049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75523,83 +56273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75824,83 +56497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76125,83 +56721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76426,83 +56945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76727,83 +57169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77028,83 +57393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77329,83 +57617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77630,83 +57841,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77931,83 +58065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78233,83 +58290,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78534,83 +58514,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78836,83 +58739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79137,83 +58963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79439,83 +59188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79740,83 +59412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80042,83 +59637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80344,83 +59862,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80646,83 +60087,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80947,83 +60311,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81248,83 +60535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81550,83 +60760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81851,83 +60984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82152,83 +61208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82454,83 +61433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82755,83 +61657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83056,83 +61881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83357,83 +62105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83658,83 +62329,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83959,83 +62553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84260,83 +62777,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84561,83 +63001,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84863,83 +63226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85164,83 +63450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85465,83 +63674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85766,83 +63898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86068,83 +64123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86369,83 +64347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86671,83 +64572,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86973,83 +64797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87274,83 +65021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87575,83 +65245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87876,83 +65469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88177,83 +65693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88478,83 +65917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88779,83 +66141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89080,83 +66365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89382,83 +66590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89683,83 +66814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89984,83 +67038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90286,83 +67263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90587,83 +67487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90888,83 +67711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91189,83 +67935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91491,83 +68160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91792,83 +68384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92093,83 +68608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92394,83 +68832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92695,83 +69056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92996,83 +69280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93297,83 +69504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93598,83 +69728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93899,83 +69952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94200,83 +70176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94502,83 +70401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94803,83 +70625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95104,83 +70849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95405,83 +71073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95707,83 +71298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96009,83 +71523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96311,83 +71748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96612,83 +71972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96913,83 +72196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97214,83 +72420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97516,83 +72645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97817,83 +72869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98118,83 +73093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98419,83 +73317,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98721,83 +73542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99022,83 +73766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99323,83 +73990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99624,83 +74214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99925,83 +74438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100226,83 +74662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100527,83 +74886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100828,83 +75110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101129,83 +75334,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101430,83 +75558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101731,83 +75782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102033,83 +76007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102335,83 +76232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102636,83 +76456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102937,83 +76680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103238,83 +76904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103540,83 +77129,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103842,83 +77354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104143,83 +77578,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104445,83 +77803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104747,83 +78028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105049,83 +78253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105350,83 +78477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105652,83 +78702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105953,83 +78926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106254,83 +79150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106555,83 +79374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106856,83 +79598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107157,83 +79822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107458,83 +80046,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107759,83 +80270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108061,83 +80495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108363,83 +80720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108665,83 +80945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108966,83 +81169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109268,83 +81394,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109569,83 +81618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109870,83 +81842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110172,83 +82067,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110473,83 +82291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110775,83 +82516,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111076,83 +82740,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111377,83 +82964,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111678,83 +83188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111979,83 +83412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112281,83 +83637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112582,83 +83861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112884,83 +84086,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113185,83 +84310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113487,83 +84535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113789,83 +84760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114090,83 +84984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114391,83 +85208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114693,83 +85433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114994,83 +85657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115295,83 +85881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115596,83 +86105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115897,83 +86329,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116198,83 +86553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116500,83 +86778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116802,83 +87003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117103,83 +87227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117405,83 +87452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117706,83 +87676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118007,83 +87900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118309,83 +88125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118610,83 +88349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118911,83 +88573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119212,83 +88797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119514,83 +89022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119815,83 +89246,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120116,83 +89470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120417,83 +89694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120719,83 +89919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121020,83 +90143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121322,83 +90368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121623,83 +90592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121924,83 +90816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122225,83 +91040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122526,83 +91264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122828,83 +91489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123129,83 +91713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123430,83 +91937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123732,83 +92162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124033,83 +92386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124334,83 +92610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124635,83 +92834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124936,83 +93058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125237,83 +93282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125538,83 +93506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125840,83 +93731,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126141,83 +93955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126442,83 +94179,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126743,83 +94403,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127044,83 +94627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127345,83 +94851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127647,83 +95076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127948,83 +95300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128249,83 +95524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128551,83 +95749,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128853,83 +95974,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129154,83 +96198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129456,83 +96423,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129757,83 +96647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130058,83 +96871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130359,83 +97095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130660,83 +97319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130961,83 +97543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131262,83 +97767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131563,83 +97991,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131864,83 +98215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132165,83 +98439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132467,83 +98664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132769,83 +98889,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133071,83 +99114,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133372,83 +99338,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133673,83 +99562,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133974,83 +99786,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134276,83 +100011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134577,83 +100235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134878,83 +100459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135179,83 +100683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135480,83 +100907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135781,83 +101131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136082,83 +101355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136383,83 +101579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136684,83 +101803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136985,83 +102027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137286,83 +102251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137587,83 +102475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137888,83 +102699,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138189,83 +102923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138490,83 +103147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138792,83 +103372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139094,83 +103597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139395,83 +103821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139696,83 +104045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139997,83 +104269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140298,83 +104493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140600,83 +104718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140901,83 +104942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141203,83 +105167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141504,83 +105391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141806,83 +105616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142107,83 +105840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142408,83 +106064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142709,83 +106288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143010,83 +106512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143311,83 +106736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143613,83 +106961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143914,83 +107185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144215,83 +107409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144516,83 +107633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144818,83 +107858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145119,83 +108082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145420,83 +108306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145722,83 +108531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146024,83 +108756,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146325,83 +108980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146626,83 +109204,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146927,83 +109428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147228,83 +109652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147530,83 +109877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147831,83 +110101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148133,83 +110326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148435,83 +110551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148736,83 +110775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149037,83 +110999,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149339,83 +111224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149640,83 +111448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149941,83 +111672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150242,83 +111896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150543,83 +112120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150844,83 +112344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151145,83 +112568,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151446,83 +112792,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151747,83 +113016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152048,83 +113240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152350,83 +113465,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152651,83 +113689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152952,83 +113913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153253,83 +114137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153555,83 +114362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153856,83 +114586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154157,83 +114810,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154459,83 +115035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154760,83 +115259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155062,83 +115484,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155363,83 +115708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155665,83 +115933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155966,83 +116157,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156267,83 +116381,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156569,83 +116606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156870,83 +116830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157171,83 +117054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157473,83 +117279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157774,83 +117503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158075,83 +117727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158376,83 +117951,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158678,83 +118176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158979,83 +118400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159280,83 +118624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159581,83 +118848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159882,83 +119072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160183,83 +119296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160484,83 +119520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160785,83 +119744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161087,83 +119969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161388,83 +120193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161689,83 +120417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161991,83 +120642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162292,83 +120866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162593,83 +121090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162894,83 +121314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163195,83 +121538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163496,83 +121762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163797,83 +121986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164099,83 +122211,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164400,83 +122435,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164701,83 +122659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165002,83 +122883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165304,83 +123108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165605,83 +123332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165907,83 +123557,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166209,83 +123782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166510,83 +124006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166811,83 +124230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167113,83 +124455,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167415,83 +124680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167716,83 +124904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168017,83 +125128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168319,83 +125353,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168620,83 +125577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168921,83 +125801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169222,83 +126025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169523,83 +126249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169824,83 +126473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170125,83 +126697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170426,83 +126921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170727,83 +127145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171029,83 +127370,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171331,83 +127595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171632,83 +127819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171933,83 +128043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172234,83 +128267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172535,83 +128491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172836,83 +128715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173137,83 +128939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173438,83 +129163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173739,83 +129387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174041,83 +129612,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174343,83 +129837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174644,83 +130061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174945,83 +130285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175246,83 +130509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175547,83 +130733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175849,83 +130958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176150,83 +131182,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176451,83 +131406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176753,83 +131631,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177054,83 +131855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177355,83 +132079,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177656,83 +132303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177958,83 +132528,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178260,83 +132753,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178561,83 +132977,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178862,83 +133201,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179163,83 +133425,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179464,83 +133649,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179766,83 +133874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180068,83 +134099,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180369,83 +134323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180670,83 +134547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180971,83 +134771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181273,83 +134996,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181574,83 +135220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181875,83 +135444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182176,83 +135668,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182478,83 +135893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182780,83 +136118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183081,83 +136342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183382,83 +136566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183683,83 +136790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183984,83 +137014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_Aux_AH_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_Aux_AH_SAV.yaml index 44ad399e0bd..179f00aaf00 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_Aux_AH_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_Aux_AH_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -401,66 +341,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -625,66 +505,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcA_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcA_SAV.yaml index 405de5d05ab..f798e988d4b 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcA_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcA_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: A - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -400,66 +340,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: A - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcB_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcB_SAV.yaml index d432edc6ca6..190712d8331 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcB_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcB_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -400,66 +340,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -623,66 +503,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml index 8cc5d35bf4f..bcab137901f 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -401,66 +341,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_GG_AS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_GG_AS_SAV_UserArgs.yaml index ff649d41944..d474372e258 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_GG_AS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_GG_AS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -70,7 +70,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 UseScaleCD: false - - 1LDSBuffer: 0 @@ -202,78 +202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: true - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml index e86b1cb8a05..121133e7ed7 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -222,81 +222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -506,81 +431,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -790,81 +640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1074,81 +849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1358,81 +1058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1642,81 +1267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1926,81 +1476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2210,81 +1685,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2494,81 +1894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2778,81 +2103,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3062,81 +2312,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3346,81 +2521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3630,81 +2730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3914,81 +2939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4198,81 +3148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4482,81 +3357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4766,81 +3566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5050,81 +3775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5334,81 +3984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5618,81 +4193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5902,81 +4402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6186,81 +4611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6470,81 +4820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6754,81 +5029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7038,81 +5238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7322,81 +5447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7606,81 +5656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7890,81 +5865,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8174,81 +6074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8458,81 +6283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8742,81 +6492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9026,81 +6701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9310,81 +6910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9594,81 +7119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9878,81 +7328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10162,81 +7537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10446,81 +7746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10730,81 +7955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11014,81 +8164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11298,81 +8373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11582,81 +8582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11866,81 +8791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12150,81 +9000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12434,81 +9209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12718,81 +9418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13002,81 +9627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13286,81 +9836,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13570,81 +10045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13854,81 +10254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14138,81 +10463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14422,81 +10672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14706,81 +10881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14990,81 +11090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15274,81 +11299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15558,81 +11508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15842,81 +11717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16126,81 +11926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16410,81 +12135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16694,81 +12344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16978,81 +12553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17262,81 +12762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17546,81 +12971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17830,81 +13180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18114,81 +13389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18398,81 +13598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18682,81 +13807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18966,81 +14016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19250,81 +14225,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19534,81 +14434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19818,81 +14643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20102,81 +14852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20386,81 +15061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20670,81 +15270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20954,81 +15479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21238,81 +15688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21522,81 +15897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21806,81 +16106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22090,81 +16315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22374,81 +16524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22658,81 +16733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22942,81 +16942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23226,81 +17151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23510,81 +17360,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23794,81 +17569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24078,81 +17778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24362,81 +17987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24646,81 +18196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24930,81 +18405,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25214,81 +18614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25498,81 +18823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25782,81 +19032,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26066,81 +19241,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26350,81 +19450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26634,81 +19659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26918,81 +19868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27202,81 +20077,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27486,81 +20286,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27770,81 +20495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28054,81 +20704,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28338,81 +20913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28622,81 +21122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28906,81 +21331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29190,81 +21540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29482,83 +21757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29783,83 +21981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30084,83 +22205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30385,83 +22429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30686,83 +22653,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30987,83 +22877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31288,83 +23101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31589,83 +23325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31890,83 +23549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32191,83 +23773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32492,83 +23997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32793,83 +24221,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33094,83 +24445,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33395,83 +24669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33696,83 +24893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33997,83 +25117,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34298,83 +25341,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34599,83 +25565,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34900,83 +25789,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35201,83 +26013,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35502,83 +26237,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35803,83 +26461,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36104,83 +26685,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36405,83 +26909,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36706,83 +27133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37008,83 +27358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37309,83 +27582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37610,83 +27806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37912,83 +28031,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38213,83 +28255,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38514,83 +28479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38815,83 +28703,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39116,83 +28927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39417,83 +29151,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39718,83 +29375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40019,83 +29599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40321,83 +29824,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40622,83 +30048,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40924,83 +30273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41225,83 +30497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41526,83 +30721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41827,83 +30945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42128,83 +31169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42429,83 +31393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42731,83 +31618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43032,83 +31842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43333,83 +32066,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43634,83 +32290,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43936,83 +32515,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44237,83 +32739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44539,83 +32964,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44841,83 +33189,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45143,83 +33414,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45444,83 +33638,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45745,83 +33862,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46046,83 +34086,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46347,83 +34310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46648,83 +34534,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46949,83 +34758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47251,83 +34983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47553,83 +35208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47854,83 +35432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48155,83 +35656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48457,83 +35881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48758,83 +36105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49059,83 +36329,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49361,83 +36554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49662,83 +36778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49963,83 +37002,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50264,83 +37226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50565,83 +37450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50866,83 +37674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51167,83 +37898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51468,83 +38122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51769,83 +38346,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52070,83 +38570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52372,83 +38795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52673,83 +39019,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52974,83 +39243,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53275,83 +39467,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53576,83 +39691,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53877,83 +39915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54178,83 +40139,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54479,83 +40363,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54780,83 +40587,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55081,83 +40811,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55383,83 +41036,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55685,83 +41261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55986,83 +41485,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56287,83 +41709,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56589,83 +41934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56890,83 +42158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57191,83 +42382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57492,83 +42606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57794,83 +42831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58095,83 +43055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58396,83 +43279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58697,83 +43503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58998,83 +43727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59299,83 +43951,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59600,83 +44175,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59901,83 +44399,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60203,83 +44624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60504,83 +44848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60806,83 +45073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61107,83 +45297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61409,83 +45522,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61710,83 +45746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62011,83 +45970,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62313,83 +46195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62614,83 +46419,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62915,83 +46643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63217,83 +46868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63518,83 +47092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63819,83 +47316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64120,83 +47540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64422,83 +47765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64724,83 +47990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65026,83 +48215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65327,83 +48439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65628,83 +48663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65929,83 +48887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66230,83 +49111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66531,83 +49335,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66833,83 +49560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67134,83 +49784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67435,83 +50008,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67736,83 +50232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68037,83 +50456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68339,83 +50681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68641,83 +50906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68942,83 +51130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69243,83 +51354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69544,83 +51578,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69845,83 +51802,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70146,83 +52026,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70447,83 +52250,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70748,83 +52474,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71049,83 +52698,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71350,83 +52922,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71651,83 +53146,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71952,83 +53370,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72253,83 +53594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72555,83 +53819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72856,83 +54043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73157,83 +54267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73458,83 +54491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73759,83 +54715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74060,83 +54939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74361,83 +55163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74662,83 +55387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74963,83 +55611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75264,83 +55835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75565,83 +56059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75866,83 +56283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76167,83 +56507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76468,83 +56731,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76769,83 +56955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77070,83 +57179,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77372,83 +57404,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77673,83 +57628,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77974,83 +57852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78275,83 +58076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78576,83 +58300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78877,83 +58524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79178,83 +58748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79479,83 +58972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79780,83 +59196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80081,83 +59420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80382,83 +59644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80683,83 +59868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80984,83 +60092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81285,83 +60316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81586,83 +60540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81887,83 +60764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82188,83 +60988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82490,83 +61213,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82791,83 +61437,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83093,83 +61662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83395,83 +61887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83696,83 +62111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83998,83 +62336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84299,83 +62560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84600,83 +62784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84901,83 +63008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85202,83 +63232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85503,83 +63456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85804,83 +63680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86105,83 +63904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86406,83 +64128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86708,83 +64353,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87009,83 +64577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87311,83 +64802,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87613,83 +65027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87914,83 +65251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88215,83 +65475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88516,83 +65699,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88818,83 +65924,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89119,83 +66148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89420,83 +66372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89721,83 +66596,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90022,83 +66820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90324,83 +67045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90625,83 +67269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90926,83 +67493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91227,83 +67717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91528,83 +67941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91830,83 +68166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92131,83 +68390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92432,83 +68614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92733,83 +68838,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93034,83 +69062,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93336,83 +69287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93637,83 +69511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93939,83 +69736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94241,83 +69961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94542,83 +70185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94843,83 +70409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95144,83 +70633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95446,83 +70858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95747,83 +71082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96049,83 +71307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96350,83 +71531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96651,83 +71755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96952,83 +71979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97253,83 +72203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97554,83 +72427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97855,83 +72651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98157,83 +72876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98458,83 +73100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98760,83 +73325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99061,83 +73549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99362,83 +73773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99663,83 +73997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99964,83 +74221,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100265,83 +74445,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100566,83 +74669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100867,83 +74893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101169,83 +75118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101470,83 +75342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101771,83 +75566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102072,83 +75790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102373,83 +76014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102674,83 +76238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102975,83 +76462,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103276,83 +76686,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103577,83 +76910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103878,83 +77134,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104180,83 +77359,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104481,83 +77583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104782,83 +77807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105083,83 +78031,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105384,83 +78255,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105685,83 +78479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105986,83 +78703,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106287,83 +78927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106589,83 +79152,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106890,83 +79376,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107192,83 +79601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107494,83 +79826,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107795,83 +80050,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108096,83 +80274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108397,83 +80498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108698,83 +80722,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109000,83 +80947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109301,83 +81171,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109602,83 +81395,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109904,83 +81620,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110206,83 +81845,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110507,83 +82069,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110808,83 +82293,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111109,83 +82517,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111410,83 +82741,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111712,83 +82966,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112014,83 +83191,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112316,83 +83416,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112618,83 +83641,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112920,83 +83866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113222,83 +84091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113523,83 +84315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113824,83 +84539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114125,83 +84763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114426,83 +84987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114727,83 +85211,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115028,83 +85435,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115330,83 +85660,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115632,83 +85885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115933,83 +86109,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116234,83 +86333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116535,83 +86557,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116837,83 +86782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117138,83 +87006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117439,83 +87230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117740,83 +87454,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118041,83 +87678,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118342,83 +87902,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118644,83 +88127,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118945,83 +88351,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119246,83 +88575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119548,83 +88800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119850,83 +89025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120151,83 +89249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120453,83 +89474,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120754,83 +89698,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121056,83 +89923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121358,83 +90148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121659,83 +90372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121960,83 +90596,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122261,83 +90820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122562,83 +91044,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122864,83 +91269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123165,83 +91493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123466,83 +91717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123768,83 +91942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124069,83 +92166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124370,83 +92390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124671,83 +92614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124973,83 +92839,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125275,83 +93064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125576,83 +93288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125877,83 +93512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126178,83 +93736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126479,83 +93960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126780,83 +94184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127081,83 +94408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127382,83 +94632,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127683,83 +94856,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127984,83 +95080,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128285,83 +95304,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128586,83 +95528,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128887,83 +95752,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129188,83 +95976,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129489,83 +96200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129790,83 +96424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130091,83 +96648,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130392,83 +96872,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130694,83 +97097,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130995,83 +97321,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131296,83 +97545,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131597,83 +97769,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131898,83 +97993,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132199,83 +98217,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132500,83 +98441,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132801,83 +98665,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133102,83 +98889,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133403,83 +99113,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133704,83 +99337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134006,83 +99562,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134307,83 +99786,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134608,83 +100010,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134910,83 +100235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135212,83 +100460,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135514,83 +100685,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135815,83 +100909,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136116,83 +101133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136417,83 +101357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136719,83 +101582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137020,83 +101806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137321,83 +102030,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137622,83 +102254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137924,83 +102479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138225,83 +102703,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138526,83 +102927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138827,83 +103151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139128,83 +103375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139430,83 +103600,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139731,83 +103824,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140033,83 +104049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140335,83 +104274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140636,83 +104498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140938,83 +104723,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141239,83 +104947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141540,83 +105171,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141841,83 +105395,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142142,83 +105619,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142443,83 +105843,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142744,83 +106067,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143046,83 +106292,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143348,83 +106517,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143649,83 +106741,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143950,83 +106965,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144251,83 +107189,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144552,83 +107413,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144853,83 +107637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145154,83 +107861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145455,83 +108085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145757,83 +108310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146058,83 +108534,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146359,83 +108758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146660,83 +108982,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146962,83 +109207,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147264,83 +109432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147565,83 +109656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147866,83 +109880,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148167,83 +110104,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148468,83 +110328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148770,83 +110553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149071,83 +110777,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149372,83 +111001,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149673,83 +111225,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149974,83 +111449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150275,83 +111673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150576,83 +111897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150878,83 +112122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151180,83 +112347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151481,83 +112571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151782,83 +112795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152083,83 +113019,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152384,83 +113243,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152685,83 +113467,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152986,83 +113691,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153288,83 +113916,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153589,83 +114140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153891,83 +114365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154192,83 +114589,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154493,83 +114813,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154794,83 +115037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155096,83 +115262,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155397,83 +115486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155698,83 +115710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155999,83 +115934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156300,83 +116158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156602,83 +116383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156903,83 +116607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157205,83 +116832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157506,83 +117056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157807,83 +117280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158108,83 +117504,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158409,83 +117728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158710,83 +117952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159011,83 +118176,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159312,83 +118400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159613,83 +118624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159915,83 +118849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160216,83 +119073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160517,83 +119297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160818,83 +119521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161119,83 +119745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161420,83 +119969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161721,83 +120193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162023,83 +120418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162324,83 +120642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162625,83 +120866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162926,83 +121090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163227,83 +121314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163528,83 +121538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163829,83 +121762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164130,83 +121986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164431,83 +122210,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164732,83 +122434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165034,83 +122659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165335,83 +122883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165636,83 +123107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165937,83 +123331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166239,83 +123556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166540,83 +123780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166841,83 +124004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167142,83 +124228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167443,83 +124452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167744,83 +124676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168045,83 +124900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168346,83 +125124,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168647,83 +125348,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168948,83 +125572,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169249,83 +125796,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169550,83 +126020,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169851,83 +126244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170152,83 +126468,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170453,83 +126692,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170754,83 +126916,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171055,83 +127140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171356,83 +127364,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171658,83 +127589,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171960,83 +127814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172262,83 +128039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172563,83 +128263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172864,83 +128487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173165,83 +128711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173466,83 +128935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173767,83 +129159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174068,83 +129383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174370,83 +129608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174671,83 +129832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174972,83 +130056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175273,83 +130280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175574,83 +130504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175875,83 +130728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176177,83 +130953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176478,83 +131177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176779,83 +131401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177080,83 +131625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177382,83 +131850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177683,83 +132074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177984,83 +132298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178285,83 +132522,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178586,83 +132746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178887,83 +132970,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179188,83 +133194,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179489,83 +133418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179790,83 +133642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180091,83 +133866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180392,83 +134090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180693,83 +134314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180994,83 +134538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181295,83 +134762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181596,83 +134986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181897,83 +135210,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182198,83 +135434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182499,83 +135658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182800,83 +135882,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183101,83 +136106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183402,83 +136330,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183703,83 +136554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184004,83 +136778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184305,83 +137002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184607,83 +137227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184909,83 +137452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185210,83 +137676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185511,83 +137900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185812,83 +138124,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186113,83 +138348,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186415,83 +138573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186716,83 +138797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187017,83 +139021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187318,83 +139245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187619,83 +139469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187920,83 +139693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188222,83 +139918,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188523,83 +140142,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188824,83 +140366,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189125,83 +140590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189427,83 +140815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189728,83 +141039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190029,83 +141263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190330,83 +141487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190632,83 +141712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190934,83 +141937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191235,83 +142161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191536,83 +142385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191837,83 +142609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192139,83 +142834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192440,83 +143058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192742,83 +143283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193044,83 +143508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193345,83 +143732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193646,83 +143956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193947,83 +144180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194248,83 +144404,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194549,83 +144628,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194851,83 +144853,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195152,83 +145077,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195454,83 +145302,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195756,83 +145527,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196058,83 +145752,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196359,83 +145976,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196660,83 +146200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196961,83 +146424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197262,83 +146648,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197564,83 +146873,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197866,83 +147098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198167,83 +147322,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198469,83 +147547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198770,83 +147771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199071,83 +147995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199372,83 +148219,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199673,83 +148443,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199974,83 +148667,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200275,83 +148891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200576,83 +149115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200877,83 +149339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201178,83 +149563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201479,83 +149787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201780,83 +150011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202082,83 +150236,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202383,83 +150460,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202684,83 +150684,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202985,83 +150908,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203286,83 +151132,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203588,83 +151357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203890,83 +151582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204191,83 +151806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204493,83 +152031,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204794,83 +152255,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205095,83 +152479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205396,83 +152703,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205697,83 +152927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205998,83 +153151,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206299,83 +153375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206600,83 +153599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206901,83 +153823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207202,83 +154047,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207503,83 +154271,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207804,83 +154495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208106,83 +154720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208408,83 +154945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208709,83 +155169,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209011,83 +155394,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209312,83 +155618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209613,83 +155842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209914,83 +156066,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210215,83 +156290,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210516,83 +156514,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210817,83 +156738,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211118,83 +156962,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211419,83 +157186,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211720,83 +157410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212021,83 +157634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212322,83 +157858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212623,83 +158082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212924,83 +158306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213226,83 +158531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213527,83 +158755,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213828,83 +158979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214129,83 +159203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214430,83 +159427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214731,83 +159651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215032,83 +159875,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215333,83 +160099,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215634,83 +160323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215936,83 +160548,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216237,83 +160772,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216538,83 +160996,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216839,83 +161220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217140,83 +161444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217442,83 +161669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217743,83 +161893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218044,83 +162117,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218345,83 +162341,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218647,83 +162566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218948,83 +162790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219250,83 +163015,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219551,83 +163239,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219852,83 +163463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220153,83 +163687,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220454,83 +163911,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220755,83 +164135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221056,83 +164359,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221357,83 +164583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221658,83 +164807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221959,83 +165031,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222260,83 +165255,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222562,83 +165480,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222864,83 +165705,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223165,83 +165929,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223466,83 +166153,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223767,83 +166377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224068,83 +166601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224370,83 +166826,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224671,83 +167050,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224972,83 +167274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225273,83 +167498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225574,83 +167722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225875,83 +167946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226176,83 +168170,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226478,83 +168395,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226780,83 +168620,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227081,83 +168844,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227382,83 +169068,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227684,83 +169293,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227985,83 +169517,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228287,83 +169742,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228588,83 +169966,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228889,83 +170190,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229190,83 +170414,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229492,83 +170639,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229794,83 +170864,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230095,83 +171088,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230396,83 +171312,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230697,83 +171536,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230999,83 +171761,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231300,83 +171985,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231602,83 +172210,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231903,83 +172434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232205,83 +172659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232506,83 +172883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232807,83 +173107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233108,83 +173331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233409,83 +173555,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233711,83 +173780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234012,83 +174004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234313,83 +174228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234614,83 +174452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234916,83 +174677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235217,83 +174901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235518,83 +175125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235820,83 +175350,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236122,83 +175575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236423,83 +175799,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236724,83 +176023,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237025,83 +176247,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237326,83 +176471,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237628,83 +176696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237929,83 +176920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238230,83 +177144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238532,83 +177369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238834,83 +177594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239135,83 +177818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239436,83 +178042,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239738,83 +178267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240040,83 +178492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240341,83 +178716,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240642,83 +178940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240943,83 +179164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241245,83 +179389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241547,83 +179614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241848,83 +179838,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242149,83 +180062,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242451,83 +180287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242753,83 +180512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243054,83 +180736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243355,83 +180960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243656,83 +181184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243957,83 +181408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244259,83 +181633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244561,83 +181858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244862,83 +182082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245164,83 +182307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245465,83 +182531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245767,83 +182756,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246068,83 +182980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246370,83 +183205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246672,83 +183430,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246973,83 +183654,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247274,83 +183878,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247575,83 +184102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247876,83 +184326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248177,83 +184550,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248479,83 +184775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248781,83 +185000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249083,83 +185225,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249385,83 +185450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249686,83 +185674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249987,83 +185898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250288,83 +186122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250589,83 +186346,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250890,83 +186570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251192,83 +186795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251493,83 +187019,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251794,83 +187243,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252095,83 +187467,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252396,83 +187691,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252697,83 +187915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HSS_BH_Bias_GG_AS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HSS_BH_Bias_GG_AS_SAV_UserArgs.yaml index 00310e47a4b..17940c5649d 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HSS_BH_Bias_GG_AS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HSS_BH_Bias_GG_AS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -70,7 +70,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 UseScaleCD: false - - 1LDSBuffer: 0 @@ -202,78 +202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: true - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HSS_BH_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HSS_BH_Bias_HAS_SAV_UserArgs.yaml index 5c285b161c3..19e0b084ff3 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HSS_BH_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_HSS_BH_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -834,83 +680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1135,83 +904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1436,83 +1128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1737,83 +1352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2039,83 +1577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2340,83 +1801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2641,83 +2025,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2942,83 +2249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3243,83 +2473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3544,83 +2697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3845,83 +2921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4146,83 +3145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4447,83 +3369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4748,83 +3593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5049,83 +3817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5350,83 +4041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5651,83 +4265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5952,83 +4489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6253,83 +4713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6555,83 +4938,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6856,83 +5162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7157,83 +5386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7458,83 +5610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7759,83 +5834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8060,83 +6058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8361,83 +6282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8662,83 +6506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8963,83 +6730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9265,83 +6955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9566,83 +7179,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9867,83 +7403,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10168,83 +7627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10469,83 +7851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10771,83 +8076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11072,83 +8300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11373,83 +8524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11674,83 +8748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11975,83 +8972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12276,83 +9196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12577,83 +9420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12878,83 +9644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13179,83 +9868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13480,83 +10092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13781,83 +10316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14082,83 +10540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14383,83 +10764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14684,83 +10988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14985,83 +11212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15286,83 +11436,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15587,83 +11660,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15888,83 +11884,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16189,83 +12108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16490,83 +12332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16791,83 +12556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17092,83 +12780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17393,83 +13004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17695,83 +13229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17996,83 +13453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18297,83 +13677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18598,83 +13901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18899,83 +14125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19200,83 +14349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19502,83 +14574,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19803,83 +14798,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20104,83 +15022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20405,83 +15246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20706,83 +15470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21007,83 +15694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21308,83 +15918,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21610,83 +16143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21911,83 +16367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22212,83 +16591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22513,83 +16815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22815,83 +17040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23116,83 +17264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23417,83 +17488,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23718,83 +17712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24019,83 +17936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24320,83 +18160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24621,83 +18384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24922,83 +18608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25223,83 +18832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25525,83 +19057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25826,83 +19281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26127,83 +19505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26429,83 +19730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26730,83 +19954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27031,83 +20178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27333,83 +20403,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27635,83 +20628,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27937,83 +20853,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28239,83 +21078,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28540,83 +21302,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28841,83 +21526,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29142,83 +21750,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29443,83 +21974,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29744,83 +22198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30045,83 +22422,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30347,83 +22647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30648,83 +22871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30950,83 +23096,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31251,83 +23320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31552,83 +23544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31853,83 +23768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32154,83 +23992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32455,83 +24216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32756,83 +24440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33057,83 +24664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33358,83 +24888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33659,83 +25112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33960,83 +25336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34261,83 +25560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34563,83 +25785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34864,83 +26009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35165,83 +26233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35466,83 +26457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35768,83 +26682,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36070,83 +26907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36371,83 +27131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36673,83 +27356,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36974,83 +27580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37275,83 +27804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37576,83 +28028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37877,83 +28252,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38178,83 +28476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38480,83 +28701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38781,83 +28925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39082,83 +29149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39383,83 +29373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39684,83 +29597,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39985,83 +29821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40286,83 +30045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40587,83 +30269,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40888,83 +30493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41189,83 +30717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41490,83 +30941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41791,83 +31165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42092,83 +31389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42393,83 +31613,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42694,83 +31837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42995,83 +32061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43296,83 +32285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43597,83 +32509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43898,83 +32733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44199,83 +32957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44500,83 +33181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44801,83 +33405,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45102,83 +33629,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45403,83 +33853,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45704,83 +34077,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46006,83 +34302,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46307,83 +34526,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46608,83 +34750,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46909,83 +34974,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47210,83 +35198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47511,83 +35422,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47812,83 +35646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48114,83 +35871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48415,83 +36095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48716,83 +36319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49017,83 +36543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49319,83 +36768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49620,83 +36992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49921,83 +37216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50222,83 +37440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50524,83 +37665,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50826,83 +37890,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51127,83 +38114,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51428,83 +38338,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51730,83 +38563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52031,83 +38787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52332,83 +39011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52634,83 +39236,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52935,83 +39460,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53236,83 +39684,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53537,83 +39908,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53838,83 +40132,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54139,83 +40356,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54440,83 +40580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54741,83 +40804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55042,83 +41028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55343,83 +41252,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55645,83 +41477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55946,83 +41701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56247,83 +41925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56548,83 +42149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56850,83 +42374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57151,83 +42598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57453,83 +42823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57754,83 +43047,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58055,83 +43271,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58356,83 +43495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58657,83 +43719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58958,83 +43943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59259,83 +44167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59560,83 +44391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59861,83 +44615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60162,83 +44839,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60463,83 +45063,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60764,83 +45287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61065,83 +45511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61367,83 +45736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61668,83 +45960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61970,83 +46185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62272,83 +46410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62573,83 +46634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62874,83 +46858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63175,83 +47082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63477,83 +47307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63778,83 +47531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64079,83 +47755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64380,83 +47979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64681,83 +48203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64982,83 +48427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65284,83 +48652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65585,83 +48876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65886,83 +49100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66187,83 +49324,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66488,83 +49548,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66789,83 +49772,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67091,83 +49997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67392,83 +50221,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67694,83 +50446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67995,83 +50670,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68296,83 +50894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68597,83 +51118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68898,83 +51342,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69199,83 +51566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69500,83 +51790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69801,83 +52014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70102,83 +52238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70403,83 +52462,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70704,83 +52686,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71005,83 +52910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71306,83 +53134,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71607,83 +53358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71908,83 +53582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72209,83 +53806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72510,83 +54030,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72811,83 +54254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73112,83 +54478,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73413,83 +54702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73714,83 +54926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74015,83 +55150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74316,83 +55374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74617,83 +55598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74918,83 +55822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75219,83 +56046,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75521,83 +56271,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75822,83 +56495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76123,83 +56719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76424,83 +56943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76725,83 +57167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77026,83 +57391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77327,83 +57615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77628,83 +57839,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77929,83 +58063,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78230,83 +58287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78531,83 +58511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78832,83 +58735,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79133,83 +58959,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79434,83 +59183,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79735,83 +59407,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80036,83 +59631,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80337,83 +59855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80638,83 +60079,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80939,83 +60303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81241,83 +60528,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81542,83 +60752,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81844,83 +60977,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82146,83 +61202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82447,83 +61426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82748,83 +61650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83049,83 +61874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83350,83 +62098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83652,83 +62323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83954,83 +62548,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84256,83 +62773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84558,83 +62998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84859,83 +63222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85161,83 +63447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85462,83 +63671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85763,83 +63895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86064,83 +64119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86366,83 +64344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86668,83 +64569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86969,83 +64793,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87271,83 +65018,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87572,83 +65242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87873,83 +65466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88175,83 +65691,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88476,83 +65915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88777,83 +66139,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89078,83 +66363,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89379,83 +66587,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89681,83 +66812,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89983,83 +67037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90284,83 +67261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90585,83 +67485,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90886,83 +67709,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91187,83 +67933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91488,83 +68157,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91789,83 +68381,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92091,83 +68606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92393,83 +68831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92694,83 +69055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92996,83 +69280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93297,83 +69504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93598,83 +69728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93899,83 +69952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94200,83 +70176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94502,83 +70401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94804,83 +70626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95106,83 +70851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95407,83 +71075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95708,83 +71299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96009,83 +71523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96311,83 +71748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96612,83 +71972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96913,83 +72196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97214,83 +72420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97516,83 +72645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97817,83 +72869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98119,83 +73094,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98420,83 +73318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98722,83 +73543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99023,83 +73767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99325,83 +73992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99626,83 +74216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99927,83 +74440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100229,83 +74665,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100530,83 +74889,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100831,83 +75113,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101132,83 +75337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101433,83 +75561,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101735,83 +75786,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102036,83 +76010,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102338,83 +76235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102639,83 +76459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102940,83 +76683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103241,83 +76907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103542,83 +77131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103843,83 +77355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104145,83 +77580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104447,83 +77805,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104748,83 +78029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105049,83 +78253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105350,83 +78477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105651,83 +78701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105952,83 +78925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106254,83 +79150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106555,83 +79374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106856,83 +79598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107158,83 +79823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107459,83 +80047,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107760,83 +80271,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108061,83 +80495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108362,83 +80719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108664,83 +80944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108965,83 +81168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109266,83 +81392,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109567,83 +81616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109868,83 +81840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110169,83 +82064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110470,83 +82288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110772,83 +82513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111074,83 +82738,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111375,83 +82962,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111676,83 +83186,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111977,83 +83410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112278,83 +83634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112579,83 +83858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112881,83 +84083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113183,83 +84308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113484,83 +84532,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113786,83 +84757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114087,83 +84981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114388,83 +85205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114689,83 +85429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114990,83 +85653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115291,83 +85877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115592,83 +86101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115894,83 +86326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116195,83 +86550,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116496,83 +86774,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116797,83 +86998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117098,83 +87222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117399,83 +87446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117701,83 +87671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118003,83 +87896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118304,83 +88120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118605,83 +88344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118906,83 +88568,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119207,83 +88792,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119508,83 +89016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119809,83 +89240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120110,83 +89464,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120412,83 +89689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120714,83 +89914,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121015,83 +90138,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121317,83 +90363,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121618,83 +90587,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121919,83 +90811,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122220,83 +91035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122521,83 +91259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122822,83 +91483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123123,83 +91707,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123424,83 +91931,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123725,83 +92155,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124026,83 +92379,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124327,83 +92603,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124628,83 +92827,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124929,83 +93051,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125230,83 +93275,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125531,83 +93499,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125832,83 +93723,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126133,83 +93947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126435,83 +94172,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126736,83 +94396,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127038,83 +94621,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127339,83 +94845,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127641,83 +95070,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127942,83 +95294,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128243,83 +95518,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128544,83 +95742,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128845,83 +95966,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129146,83 +96190,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129447,83 +96414,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129749,83 +96639,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130050,83 +96863,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130351,83 +97087,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130653,83 +97312,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130954,83 +97536,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131255,83 +97760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131557,83 +97985,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131858,83 +98209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132160,83 +98434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132461,83 +98658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132762,83 +98882,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133063,83 +99106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133365,83 +99331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133666,83 +99555,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133967,83 +99779,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134268,83 +100003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134570,83 +100228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134871,83 +100452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135172,83 +100676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135473,83 +100900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135775,83 +101125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136076,83 +101349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136377,83 +101573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136679,83 +101798,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136980,83 +102022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137281,83 +102246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137582,83 +102470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137883,83 +102694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138184,83 +102918,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138486,83 +103143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138787,83 +103367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139089,83 +103592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139390,83 +103816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139692,83 +104041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139993,83 +104265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140294,83 +104489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140596,83 +104714,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140897,83 +104938,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141198,83 +105162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141499,83 +105386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141800,83 +105610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142102,83 +105835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142404,83 +106060,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142706,83 +106285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143008,83 +106510,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143309,83 +106734,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143610,83 +106958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143911,83 +107182,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144212,83 +107406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144513,83 +107630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144815,83 +107855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145116,83 +108079,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145417,83 +108303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145718,83 +108527,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146020,83 +108752,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146321,83 +108976,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146622,83 +109200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146923,83 +109424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147224,83 +109648,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147526,83 +109873,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147827,83 +110097,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148128,83 +110321,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148430,83 +110546,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148731,83 +110770,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149033,83 +110995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149335,83 +111220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149636,83 +111444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149938,83 +111669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150240,83 +111894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150541,83 +112118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150842,83 +112342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151143,83 +112566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151444,83 +112790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151745,83 +113014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152047,83 +113239,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152348,83 +113463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152649,83 +113687,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152950,83 +113911,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153251,83 +114135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153552,83 +114359,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153854,83 +114584,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154155,83 +114808,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154456,83 +115032,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154758,83 +115257,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155059,83 +115481,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155360,83 +115705,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155661,83 +115929,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155962,83 +116153,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156263,83 +116377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156565,83 +116602,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156867,83 +116827,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157169,83 +117052,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157471,83 +117277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157772,83 +117501,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158073,83 +117725,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158374,83 +117949,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158675,83 +118173,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158976,83 +118397,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159277,83 +118621,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159579,83 +118846,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159880,83 +119070,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160181,83 +119294,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160482,83 +119518,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160783,83 +119742,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161085,83 +119967,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161386,83 +120191,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161687,83 +120415,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161988,83 +120639,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162289,83 +120863,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162591,83 +121088,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162893,83 +121313,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163195,83 +121538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163496,83 +121762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163797,83 +121986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164098,83 +122210,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164399,83 +122434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164700,83 +122658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165002,83 +122883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165303,83 +123107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165604,83 +123331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165905,83 +123555,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166206,83 +123779,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166507,83 +124003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166808,83 +124227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167109,83 +124451,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167411,83 +124676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167713,83 +124901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168014,83 +125125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168316,83 +125350,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168617,83 +125574,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168919,83 +125799,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169221,83 +126024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169522,83 +126248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169823,83 +126472,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170125,83 +126697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170427,83 +126922,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170729,83 +127147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171030,83 +127371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171332,83 +127596,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171634,83 +127821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171936,83 +128046,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172237,83 +128270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172538,83 +128494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172840,83 +128719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173141,83 +128943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173442,83 +129167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173744,83 +129392,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174045,83 +129616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174346,83 +129840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174647,83 +130064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174948,83 +130288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175249,83 +130512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175550,83 +130736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175851,83 +130960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176152,83 +131184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176454,83 +131409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176756,83 +131634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177057,83 +131858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177359,83 +132083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177661,83 +132308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177963,83 +132533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178264,83 +132757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178565,83 +132981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178866,83 +133205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179167,83 +133429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179468,83 +133653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179769,83 +133877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180070,83 +134101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180371,83 +134325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180672,83 +134549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180973,83 +134773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181275,83 +134998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181577,83 +135223,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181878,83 +135447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182179,83 +135671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182480,83 +135895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182781,83 +136119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183082,83 +136343,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183383,83 +136567,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183684,83 +136791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183986,83 +137016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184287,83 +137240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184588,83 +137464,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184889,83 +137688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185190,83 +137912,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185492,83 +138137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185793,83 +138361,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186094,83 +138585,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186395,83 +138809,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186696,83 +139033,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186997,83 +139257,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187298,83 +139481,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187600,83 +139706,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187901,83 +139930,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188203,83 +140155,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188505,83 +140380,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188806,83 +140604,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189107,83 +140828,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189408,83 +141052,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189709,83 +141276,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190010,83 +141500,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190311,83 +141724,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190613,83 +141949,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190914,83 +142173,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191216,83 +142398,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191518,83 +142623,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191819,83 +142847,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192121,83 +143072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192423,83 +143297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192724,83 +143521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193025,83 +143745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193326,83 +143969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193627,83 +144193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193928,83 +144417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194230,83 +144642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194531,83 +144866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194832,83 +145090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195133,83 +145314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195434,83 +145538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195736,83 +145763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196038,83 +145988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196339,83 +146212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196640,83 +146436,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196941,83 +146660,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197242,83 +146884,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197543,83 +147108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197844,83 +147332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198145,83 +147556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198446,83 +147780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198748,83 +148005,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199049,83 +148229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199350,83 +148453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199651,83 +148677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199952,83 +148901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200254,83 +149126,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200555,83 +149350,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200857,83 +149575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201159,83 +149800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201461,83 +150025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201762,83 +150249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202064,83 +150474,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202365,83 +150698,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202666,83 +150922,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202968,83 +151147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203269,83 +151371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203570,83 +151595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203872,83 +151820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204173,83 +152044,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204475,83 +152269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204776,83 +152493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205077,83 +152717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205378,83 +152941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205679,83 +153165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205980,83 +153389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206282,83 +153614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206584,83 +153839,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206885,83 +154063,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207186,83 +154287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207487,83 +154511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207788,83 +154735,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208089,83 +154959,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208391,83 +155184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208692,83 +155408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208993,83 +155632,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209295,83 +155857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209596,83 +156081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209898,83 +156306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210199,83 +156530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210500,83 +156754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210801,83 +156978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211103,83 +157203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211405,83 +157428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211706,83 +157652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212008,83 +157877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212309,83 +158101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212611,83 +158326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212912,83 +158550,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213213,83 +158774,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213515,83 +158999,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213817,83 +159224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214118,83 +159448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214419,83 +159672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214720,83 +159896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215021,83 +160120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215322,83 +160344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215623,83 +160568,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215925,83 +160793,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216226,83 +161017,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216527,83 +161241,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216828,83 +161465,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217129,83 +161689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217430,83 +161913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217731,83 +162137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218033,83 +162362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218334,83 +162586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218635,83 +162810,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218936,83 +163034,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219237,83 +163258,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219538,83 +163482,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219839,83 +163706,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220140,83 +163930,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220441,83 +164154,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220742,83 +164378,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221043,83 +164602,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221345,83 +164827,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221646,83 +165051,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221947,83 +165275,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222249,83 +165500,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222550,83 +165724,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222852,83 +165949,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223153,83 +166173,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223454,83 +166397,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223755,83 +166621,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224057,83 +166846,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224359,83 +167071,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224660,83 +167295,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224962,83 +167520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225264,83 +167745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225565,83 +167969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225867,83 +168194,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226168,83 +168418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226469,83 +168642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226770,83 +168866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227071,83 +169090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227372,83 +169314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227673,83 +169538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227974,83 +169762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228276,83 +169987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228577,83 +170211,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228878,83 +170435,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229179,83 +170659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229481,83 +170884,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229782,83 +171108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230083,83 +171332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230384,83 +171556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230685,83 +171780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230986,83 +172004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231288,83 +172229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231589,83 +172453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231891,83 +172678,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232192,83 +172902,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232493,83 +173126,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232794,83 +173350,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233096,83 +173575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233397,83 +173799,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233698,83 +174023,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233999,83 +174247,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234300,83 +174471,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234601,83 +174695,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234902,83 +174919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235203,83 +175143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235504,83 +175367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235805,83 +175591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236106,83 +175815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236407,83 +176039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236708,83 +176263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237009,83 +176487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237310,83 +176711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237611,83 +176935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237912,83 +177159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238213,83 +177383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238514,83 +177607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238816,83 +177832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239117,83 +178056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239418,83 +178280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239719,83 +178504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240021,83 +178729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240322,83 +178953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240623,83 +179177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240924,83 +179401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241225,83 +179625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241527,83 +179850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241828,83 +180074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242129,83 +180298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242430,83 +180522,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242732,83 +180747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243034,83 +180972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243335,83 +181196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243636,83 +181420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243937,83 +181644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244238,83 +181868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244539,83 +182092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244840,83 +182316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245141,83 +182540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245442,83 +182764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245743,83 +182988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246044,83 +183212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246345,83 +183436,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246647,83 +183661,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246948,83 +183885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247250,83 +184110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247552,83 +184335,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247854,83 +184560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248155,83 +184784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248456,83 +185008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248757,83 +185232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249059,83 +185457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249360,83 +185681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249662,83 +185906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249964,83 +186131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250265,83 +186355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250566,83 +186579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250868,83 +186804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251169,83 +187028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251470,83 +187252,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251771,83 +187476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252072,83 +187700,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252373,83 +187924,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252675,83 +188149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252977,83 +188374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253278,83 +188598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253579,83 +188822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253880,83 +189046,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254182,83 +189271,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254484,83 +189496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254785,83 +189720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255086,83 +189944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255387,83 +190168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255689,83 +190393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255990,83 +190617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256291,83 +190841,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256592,83 +191065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256894,83 +191290,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257195,83 +191514,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257497,83 +191739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257799,83 +191964,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258100,83 +192188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258402,83 +192413,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8BH_AI_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8BH_AI_SAV_UserArgs.yaml index 578cc4b30c1..0d2498d5346 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8BH_AI_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8BH_AI_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -70,7 +70,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 UseScaleCD: false - - 1LDSBuffer: 0 @@ -206,78 +206,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -465,78 +393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -724,78 +580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -983,78 +767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1242,78 +954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8I8S_BH_HAS_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8I8S_BH_HAS_Bias_HAS_SAV_UserArgs.yaml index 252de3f1c56..5f249c1b826 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8I8S_BH_HAS_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8I8S_BH_HAS_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -222,83 +222,6 @@ PrefetchGlobalRead: 0 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 0 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8II_BH_AI_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8II_BH_AI_SAV_UserArgs.yaml index 5a9dbd0c553..e205e061dd9 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8II_BH_AI_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8II_BH_AI_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -70,7 +70,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 UseScaleCD: false - - 1LDSBuffer: 0 @@ -207,78 +207,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -461,77 +389,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -713,77 +570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -965,77 +751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1217,77 +932,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1469,77 +1113,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8II_BH_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8II_BH_UserArgs.yaml index 999d8177576..dffda8b74a0 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8II_BH_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8II_BH_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -233,83 +233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -535,83 +458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -837,83 +683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1139,83 +908,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1441,83 +1133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1743,83 +1358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2045,83 +1583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2347,83 +1808,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2649,83 +2033,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2951,83 +2258,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3253,83 +2483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3555,83 +2708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3857,83 +2933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4159,83 +3158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4461,83 +3383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4763,83 +3608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5065,83 +3833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5367,83 +4058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5669,83 +4283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5971,83 +4508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6273,83 +4733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6575,83 +4958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6877,83 +5183,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7179,83 +5408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7481,83 +5633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7783,83 +5858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8085,83 +6083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8387,83 +6308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8689,83 +6533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8991,83 +6758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9293,83 +6983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9595,83 +7208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9897,83 +7433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10199,83 +7658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10501,83 +7883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10803,83 +8108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11105,83 +8333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11407,83 +8558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11709,83 +8783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12011,83 +9008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12313,83 +9233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12615,83 +9458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12917,83 +9683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13219,83 +9908,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13521,83 +10133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 2 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13823,83 +10358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14125,83 +10583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14427,83 +10808,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 2 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14729,83 +11033,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15031,83 +11258,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 2 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8IS_BH_HAS_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8IS_BH_HAS_Bias_HAS_SAV_UserArgs.yaml index 2a018392863..64f52e2ecdb 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8IS_BH_HAS_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8IS_BH_HAS_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -222,83 +222,6 @@ PrefetchGlobalRead: 0 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 0 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8_BH_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8_BH_UserArgs.yaml index e1d1308daee..6cc02a4ee2f 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8_BH_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_I8_BH_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -233,83 +233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -535,83 +458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -837,83 +683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1139,83 +908,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1441,83 +1133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1743,83 +1358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2045,83 +1583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2347,83 +1808,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2649,83 +2033,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2951,83 +2258,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3253,83 +2483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3555,83 +2708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3857,83 +2933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4159,83 +3158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4461,83 +3383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4763,83 +3608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5065,83 +3833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5367,83 +4058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5669,83 +4283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5971,83 +4508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6273,83 +4733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6575,83 +4958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6877,83 +5183,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7179,83 +5408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7481,83 +5633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7783,83 +5858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8085,83 +6083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8387,83 +6308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8689,83 +6533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8991,83 +6758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9293,83 +6983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9595,83 +7208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9897,83 +7433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10199,83 +7658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10501,83 +7883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 2 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10803,83 +8108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11105,83 +8333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 2 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_Aux_A_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_Aux_A_SAV_UserArgs.yaml index 70323853321..2f0b92c8f91 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_Aux_A_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_Aux_A_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -67,7 +67,7 @@ UseE: true UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 - - 1LDSBuffer: 1 ActivationAlt: false @@ -195,74 +195,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_BiasSrcB_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_BiasSrcB_SAV_UserArgs.yaml index 696953798c5..42f26f7991d 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_BiasSrcB_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_BiasSrcB_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -67,7 +67,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 - - 1LDSBuffer: 1 ActivationAlt: false @@ -195,74 +195,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 1 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_BiasSrcD_Grad_A_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_BiasSrcD_Grad_A_SAV_UserArgs.yaml index ac700d02560..0de72392a0a 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_BiasSrcD_Grad_A_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_SB_Bias_BiasSrcD_Grad_A_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -67,7 +67,7 @@ UseE: true UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 - - 1LDSBuffer: 1 ActivationAlt: false @@ -195,74 +195,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 1 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_S_B_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_S_B_Bias_HAS_SAV_UserArgs.yaml index daadd5a687f..bf817ec19b5 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_S_B_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_S_B_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -834,83 +680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1135,83 +904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1436,83 +1128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1737,83 +1352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2038,83 +1576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2339,83 +1800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2640,83 +2024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2941,83 +2248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3242,83 +2472,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3543,83 +2696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3844,83 +2920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4145,83 +3144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4446,83 +3368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4748,83 +3593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5049,83 +3817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5350,83 +4041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5651,83 +4265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5952,83 +4489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6253,83 +4713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6554,83 +4937,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6855,83 +5161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7157,83 +5386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7458,83 +5610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7759,83 +5834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8061,83 +6059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8363,83 +6284,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8664,83 +6508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8965,83 +6732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9266,83 +6956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9567,83 +7180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9869,83 +7405,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10171,83 +7630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10473,83 +7855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10774,83 +8079,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11076,83 +8304,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11377,83 +8528,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11678,83 +8752,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11979,83 +8976,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12281,83 +9201,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12583,83 +9426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12885,83 +9651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13186,83 +9875,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13487,83 +10099,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13789,83 +10324,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14091,83 +10549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14392,83 +10773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14693,83 +10997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14994,83 +11221,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15296,83 +11446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15598,83 +11671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15899,83 +11895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16200,83 +12119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16501,83 +12343,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16802,83 +12567,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17103,83 +12791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17404,83 +13015,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17705,83 +13239,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18006,83 +13463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18307,83 +13687,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18608,83 +13911,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18909,83 +14135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19210,83 +14359,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19511,83 +14583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19812,83 +14807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20113,83 +15031,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20415,83 +15256,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20716,83 +15480,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21017,83 +15704,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21318,83 +15928,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21619,83 +16152,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21920,83 +16376,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22221,83 +16600,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22522,83 +16824,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22823,83 +17048,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23125,83 +17273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23427,83 +17498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23728,83 +17722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24029,83 +17946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24330,83 +18170,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24631,83 +18394,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24932,83 +18618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25233,83 +18842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25534,83 +19066,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25835,83 +19290,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26137,83 +19515,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26438,83 +19739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26739,83 +19963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27041,83 +20188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27343,83 +20413,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27644,83 +20637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27945,83 +20861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28246,83 +21085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28548,83 +21310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28850,83 +21535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29151,83 +21759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29453,83 +21984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29755,83 +22209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30056,83 +22433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30357,83 +22657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30658,83 +22881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30959,83 +23105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31260,83 +23329,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31562,83 +23554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31863,83 +23778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32164,83 +24002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32466,83 +24227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32767,83 +24451,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33068,83 +24675,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33369,83 +24899,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33670,83 +25123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33971,83 +25347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34272,83 +25571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34573,83 +25795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34874,83 +26019,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35176,83 +26244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35477,83 +26468,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35778,83 +26692,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36079,83 +26916,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36380,83 +27140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36682,83 +27365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36984,83 +27590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37285,83 +27814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37586,83 +28038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37887,83 +28262,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38189,83 +28487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38491,83 +28712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38793,83 +28937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39095,83 +29162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39396,83 +29386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39697,83 +29610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39998,83 +29834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40299,83 +30058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40600,83 +30282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40901,83 +30506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41203,83 +30731,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41504,83 +30955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41806,83 +31180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42107,83 +31404,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42409,83 +31629,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42710,83 +31853,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43011,83 +32077,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43312,83 +32301,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43613,83 +32525,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43915,83 +32750,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44217,83 +32975,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44518,83 +33199,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44819,83 +33423,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45120,83 +33647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45421,83 +33871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45722,83 +34095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46023,83 +34319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46324,83 +34543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46625,83 +34767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46926,83 +34991,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47227,83 +35215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47528,83 +35439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47830,83 +35664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48131,83 +35888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48432,83 +36112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48734,83 +36337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49035,83 +36561,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49336,83 +36785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49637,83 +37009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49938,83 +37233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50239,83 +37457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50540,83 +37681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50841,83 +37905,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51142,83 +38129,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51443,83 +38353,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51744,83 +38577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52045,83 +38801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52346,83 +39025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52647,83 +39249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52948,83 +39473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53250,83 +39698,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53551,83 +39922,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53853,83 +40147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54154,83 +40371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54456,83 +40596,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54757,83 +40820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55058,83 +41044,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55360,83 +41269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55662,83 +41494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55963,83 +41718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56264,83 +41942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56565,83 +42166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56866,83 +42390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57168,83 +42615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57469,83 +42839,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57770,83 +43063,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58071,83 +43287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58372,83 +43511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58674,83 +43736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58975,83 +43960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59276,83 +44184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59577,83 +44408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59878,83 +44632,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60179,83 +44856,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60481,83 +45081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60783,83 +45306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61085,83 +45531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61387,83 +45756,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61689,83 +45981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61991,83 +46206,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62293,83 +46431,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62595,83 +46656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62896,83 +46880,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63197,83 +47104,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63498,83 +47328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63799,83 +47552,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64100,83 +47776,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64401,83 +48000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64702,83 +48224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65004,83 +48449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65305,83 +48673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65606,83 +48897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65907,83 +49121,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66208,83 +49345,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66510,83 +49570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66812,83 +49795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67113,83 +50019,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67414,83 +50243,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67715,83 +50467,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68017,83 +50692,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68318,83 +50916,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68620,83 +51141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68921,83 +51365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69223,83 +51590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69524,83 +51814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69826,83 +52039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70127,83 +52263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70428,83 +52487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70729,83 +52711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71030,83 +52935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71331,83 +53159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71633,83 +53384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71934,83 +53608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72236,83 +53833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72538,83 +54058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72839,83 +54282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73141,83 +54507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73443,83 +54732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73744,83 +54956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74045,83 +55180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74346,83 +55404,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74648,83 +55629,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74949,83 +55853,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75250,83 +56077,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75551,83 +56301,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75852,83 +56525,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76153,83 +56749,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76454,83 +56973,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76756,83 +57198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77057,83 +57422,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77358,83 +57646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77659,83 +57870,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77960,83 +58094,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78261,83 +58318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78562,83 +58542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78863,83 +58766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79165,83 +58991,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79466,83 +59215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79768,83 +59440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80070,83 +59665,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80371,83 +59889,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80672,83 +60113,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80973,83 +60337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81274,83 +60561,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81575,83 +60785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81876,83 +61009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82177,83 +61233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82478,83 +61457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82779,83 +61681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83080,83 +61905,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83381,83 +62129,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83682,83 +62353,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83983,83 +62577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84284,83 +62801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84585,83 +63025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84886,83 +63249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85188,83 +63474,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85489,83 +63698,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85790,83 +63922,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86091,83 +64146,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86392,83 +64370,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86693,83 +64594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86994,83 +64818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87296,83 +65043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87597,83 +65267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87898,83 +65491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88200,83 +65716,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88501,83 +65940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88802,83 +66164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89103,83 +66388,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89404,83 +66612,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89706,83 +66837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90007,83 +67061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90308,83 +67285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90610,83 +67510,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90912,83 +67735,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91214,83 +67960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91515,83 +68184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91817,83 +68409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92118,83 +68633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92420,83 +68858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92721,83 +69082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93022,83 +69306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93323,83 +69530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93624,83 +69754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93925,83 +69978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94226,83 +70202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94527,83 +70426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94828,83 +70650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95129,83 +70874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95430,83 +71098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95732,83 +71323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96034,83 +71548,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96336,83 +71773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96637,83 +71997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96938,83 +72221,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97239,83 +72445,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97540,83 +72669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97842,83 +72894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98143,83 +73118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98444,83 +73342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98746,83 +73567,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99047,83 +73791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99349,83 +74016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99650,83 +74240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99952,83 +74465,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100254,83 +74690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100555,83 +74914,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100856,83 +75138,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101158,83 +75363,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101460,83 +75588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101762,83 +75813,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102063,83 +76037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102364,83 +76261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102665,83 +76485,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102966,83 +76709,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103268,83 +76934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103569,83 +77158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103871,83 +77383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104173,83 +77608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104474,83 +77832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104775,83 +78056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105076,83 +78280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105377,83 +78504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105678,83 +78728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105979,83 +78952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106281,83 +79177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106583,83 +79402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106884,83 +79626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107186,83 +79851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107487,83 +80075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107788,83 +80299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108089,83 +80523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108390,83 +80747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108691,83 +80971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108993,83 +81196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109295,83 +81421,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109596,83 +81645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109897,83 +81869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110198,83 +82093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110499,83 +82317,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110800,83 +82541,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111102,83 +82766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111404,83 +82991,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111706,83 +83216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112008,83 +83441,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112310,83 +83666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112612,83 +83891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112913,83 +84115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113214,83 +84339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113515,83 +84563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113816,83 +84787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114118,83 +85012,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114419,83 +85236,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114720,83 +85460,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115021,83 +85684,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115322,83 +85908,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115623,83 +86132,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115924,83 +86356,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116225,83 +86580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116527,83 +86805,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116828,83 +87029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117129,83 +87253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117430,83 +87477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117731,83 +87701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118032,83 +87925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118333,83 +88149,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118635,83 +88374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118936,83 +88598,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119237,83 +88822,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119538,83 +89046,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119839,83 +89270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120140,83 +89494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120441,83 +89718,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120742,83 +89942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121043,83 +90166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121345,83 +90391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121646,83 +90615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121948,83 +90840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122250,83 +91065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122551,83 +91289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122852,83 +91513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123153,83 +91737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123454,83 +91961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123755,83 +92185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124056,83 +92409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124357,83 +92633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124659,83 +92858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124961,83 +93083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125262,83 +93307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125563,83 +93531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125864,83 +93755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126166,83 +93980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126468,83 +94205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126770,83 +94430,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127071,83 +94654,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127372,83 +94878,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127673,83 +95102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127974,83 +95326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128275,83 +95550,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128576,83 +95774,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128877,83 +95998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129178,83 +96222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129479,83 +96446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129780,83 +96670,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130081,83 +96894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130383,83 +97119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130684,83 +97343,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130985,83 +97567,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131286,83 +97791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131587,83 +98015,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131889,83 +98240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132190,83 +98464,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132491,83 +98688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132792,83 +98912,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133094,83 +99137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133395,83 +99361,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133696,83 +99585,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133998,83 +99810,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134299,83 +100034,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134600,83 +100258,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134902,83 +100483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135204,83 +100708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135506,83 +100933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135808,83 +101158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136109,83 +101382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136410,83 +101606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136711,83 +101830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137012,83 +102054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137314,83 +102279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137615,83 +102503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137916,83 +102727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138217,83 +102951,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138518,83 +103175,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138819,83 +103399,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139120,83 +103623,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139421,83 +103847,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139722,83 +104071,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140023,83 +104295,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140324,83 +104519,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140625,83 +104743,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140926,83 +104967,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141227,83 +105191,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141529,83 +105416,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141831,83 +105641,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142132,83 +105865,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142433,83 +106089,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142735,83 +106314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143037,83 +106539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143339,83 +106764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143640,83 +106988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143941,83 +107212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144242,83 +107436,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144544,83 +107661,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144846,83 +107886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145147,83 +108110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145448,83 +108334,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145749,83 +108558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146050,83 +108782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146352,83 +109007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146653,83 +109231,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146954,83 +109455,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147256,83 +109680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147557,83 +109904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147858,83 +110128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148159,83 +110352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148460,83 +110576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148761,83 +110800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149062,83 +111024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149363,83 +111248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149664,83 +111472,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149965,83 +111696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150266,83 +111920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150567,83 +112144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150869,83 +112369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151170,83 +112593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151472,83 +112818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151774,83 +113043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152075,83 +113267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152376,83 +113491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152678,83 +113716,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152979,83 +113940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153280,83 +114164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153582,83 +114389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153884,83 +114614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154185,83 +114838,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154486,83 +115062,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154788,83 +115287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155089,83 +115511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155390,83 +115735,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155691,83 +115959,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155992,83 +116183,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156293,83 +116407,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156595,83 +116632,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156896,83 +116856,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157197,83 +117080,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157499,83 +117305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157800,83 +117529,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158101,83 +117753,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158403,83 +117978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158704,83 +118202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159005,83 +118426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159307,83 +118651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159608,83 +118875,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159909,83 +119099,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160210,83 +119323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160512,83 +119548,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160814,83 +119773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161116,83 +119998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161417,83 +120222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161718,83 +120446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162020,83 +120671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162321,83 +120895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162623,83 +121120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162924,83 +121344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163225,83 +121568,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163526,83 +121792,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163827,83 +122016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164128,83 +122240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164430,83 +122465,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164732,83 +122690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165033,83 +122914,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165334,83 +123138,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165635,83 +123362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165936,83 +123586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166237,83 +123810,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166538,83 +124034,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166840,83 +124259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167141,83 +124483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167442,83 +124707,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167743,83 +124931,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168045,83 +125156,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168346,83 +125380,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168647,83 +125604,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168949,83 +125829,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169250,83 +126053,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169551,83 +126277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169853,83 +126502,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170155,83 +126727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170457,83 +126952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170758,83 +127176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171059,83 +127400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171360,83 +127624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171661,83 +127848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171962,83 +128072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172263,83 +128296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172564,83 +128520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172865,83 +128744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173167,83 +128969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173468,83 +129193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173769,83 +129417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174071,83 +129642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174372,83 +129866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174674,83 +130091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174975,83 +130315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175276,83 +130539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175577,83 +130763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175879,83 +130988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176180,83 +131212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176481,83 +131436,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176782,83 +131660,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177083,83 +131884,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177384,83 +132108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177685,83 +132332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177986,83 +132556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178287,83 +132780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178588,83 +133004,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178890,83 +133229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179191,83 +133453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179493,83 +133678,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179794,83 +133902,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180095,83 +134126,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180396,83 +134350,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180697,83 +134574,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180999,83 +134799,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181300,83 +135023,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181601,83 +135247,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181902,83 +135471,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182204,83 +135696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182505,83 +135920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182807,83 +136145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183109,83 +136370,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183411,83 +136595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183712,83 +136819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184013,83 +137043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184314,83 +137267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184615,83 +137491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184916,83 +137715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185218,83 +137940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185520,83 +138165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185821,83 +138389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186122,83 +138613,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186423,83 +138837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186724,83 +139061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187026,83 +139286,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187328,83 +139511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187629,83 +139735,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187930,83 +139959,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188231,83 +140183,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188533,83 +140408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188835,83 +140633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189136,83 +140857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189438,83 +141082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189739,83 +141306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190040,83 +141530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190341,83 +141754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190643,83 +141979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190944,83 +142203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191246,83 +142428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191547,83 +142652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191848,83 +142876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192150,83 +143101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192452,83 +143326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192754,83 +143551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193055,83 +143775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193356,83 +143999,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193658,83 +144224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193959,83 +144448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194261,83 +144673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194563,83 +144898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194864,83 +145122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195166,83 +145347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195467,83 +145571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195768,83 +145795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196069,83 +146019,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196371,83 +146244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196672,83 +146468,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196974,83 +146693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197275,83 +146917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197576,83 +147141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197877,83 +147365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198178,83 +147589,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198480,83 +147814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198782,83 +148039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199083,83 +148263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199384,83 +148487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199685,83 +148711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199986,83 +148935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200288,83 +149160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200590,83 +149385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200891,83 +149609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201193,83 +149834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201494,83 +150058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201795,83 +150282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202096,83 +150506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202397,83 +150730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202698,83 +150954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202999,83 +151178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203301,83 +151403,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203602,83 +151627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203903,83 +151851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204205,83 +152076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204506,83 +152300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204807,83 +152524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205108,83 +152748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205409,83 +152972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205711,83 +153197,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206012,83 +153421,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206313,83 +153645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206614,83 +153869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206915,83 +154093,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207216,83 +154317,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207517,83 +154541,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207818,83 +154765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208119,83 +154989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208420,83 +155213,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208722,83 +155438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209023,83 +155662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209324,83 +155886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209625,83 +156110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209926,83 +156334,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210228,83 +156559,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210530,83 +156784,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210831,83 +157008,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211132,83 +157232,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211433,83 +157456,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211734,83 +157680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212035,83 +157904,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212336,83 +158128,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212637,83 +158352,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212938,83 +158576,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213239,83 +158800,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213540,83 +159024,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213842,83 +159249,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214143,83 +159473,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214444,83 +159697,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214745,83 +159921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215046,83 +160145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215347,83 +160369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215649,83 +160594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215951,83 +160819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216253,83 +161044,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216554,83 +161268,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216855,83 +161492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217156,83 +161716,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217458,83 +161941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217759,83 +162165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218061,83 +162390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218363,83 +162615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218664,83 +162839,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218965,83 +163063,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219266,83 +163287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219567,83 +163511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219868,83 +163735,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220170,83 +163960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220471,83 +164184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220772,83 +164408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221074,83 +164633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221376,83 +164858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221677,83 +165082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221978,83 +165306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222279,83 +165530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222580,83 +165754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222881,83 +165978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223182,83 +166202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223484,83 +166427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223785,83 +166651,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224086,83 +166875,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224387,83 +167099,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224688,83 +167323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224989,83 +167547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225291,83 +167772,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225593,83 +167997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225895,83 +168222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226197,83 +168447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226498,83 +168671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226799,83 +168895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227100,83 +169119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227402,83 +169344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227703,83 +169568,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228004,83 +169792,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228305,83 +170016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228606,83 +170240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228908,83 +170465,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229209,83 +170689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229510,83 +170913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229811,83 +171137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230113,83 +171362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230414,83 +171586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230715,83 +171810,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231016,83 +172034,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231317,83 +172258,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231618,83 +172482,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231920,83 +172707,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232221,83 +172931,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232523,83 +173156,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232824,83 +173380,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233126,83 +173605,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233427,83 +173829,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233729,83 +174054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234031,83 +174279,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234332,83 +174503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234634,83 +174728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234935,83 +174952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235236,83 +175176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235537,83 +175400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235838,83 +175624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236139,83 +175848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236440,83 +176072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236741,83 +176296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237042,83 +176520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237344,83 +176745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237645,83 +176969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237947,83 +177194,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238248,83 +177418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238550,83 +177643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238851,83 +177867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239152,83 +178091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239453,83 +178315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239754,83 +178539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240055,83 +178763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240356,83 +178987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240658,83 +179212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240960,83 +179437,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241262,83 +179662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241563,83 +179886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241865,83 +180111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242166,83 +180335,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242467,83 +180559,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242768,83 +180783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243069,83 +181007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243370,83 +181231,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243672,83 +181456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243974,83 +181681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244276,83 +181906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244577,83 +182130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244878,83 +182354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245180,83 +182579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245481,83 +182803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245782,83 +183027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246083,83 +183251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246384,83 +183475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246685,83 +183699,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246986,83 +183923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247288,83 +184148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247589,83 +184372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247890,83 +184596,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248191,83 +184820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248492,83 +185044,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248793,83 +185268,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249094,83 +185492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249396,83 +185717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249697,83 +185941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249999,83 +186166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250301,83 +186391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250602,83 +186615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250903,83 +186839,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251204,83 +187063,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251506,83 +187288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251807,83 +187512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252108,83 +187736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252409,83 +187960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252711,83 +188185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253013,83 +188410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253315,83 +188635,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253616,83 +188859,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253917,83 +189083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254218,83 +189307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254519,83 +189531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254820,83 +189755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255121,83 +189979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255422,83 +190203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255723,83 +190427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256024,83 +190651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256326,83 +190876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256627,83 +191100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256928,83 +191324,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257230,83 +191549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257531,83 +191773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257832,83 +191997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258133,83 +192221,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258434,83 +192445,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258735,83 +192669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -259036,83 +192893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -259338,83 +193118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -259640,83 +193343,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -259942,83 +193568,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -260244,83 +193793,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -260545,83 +194017,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -260846,83 +194241,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -261148,83 +194466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -261449,83 +194690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -261750,83 +194914,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -262051,83 +195138,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -262352,83 +195362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -262654,83 +195587,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -262956,83 +195812,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -263258,83 +196037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -263559,83 +196261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -263861,83 +196486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -264162,83 +196710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -264464,83 +196935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -264765,83 +197159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -265067,83 +197384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -265369,83 +197609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -265670,83 +197833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -265972,83 +198058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -266274,83 +198283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -266576,83 +198508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -266877,83 +198732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -267179,83 +198957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -267480,83 +199181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -267781,83 +199405,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -268082,83 +199629,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -268384,83 +199854,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -268686,83 +200079,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -268987,83 +200303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -269289,83 +200528,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -269591,83 +200753,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -269892,83 +200977,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -270194,83 +201202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -270495,83 +201426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -270796,83 +201650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -271097,83 +201874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -271399,83 +202099,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -271700,83 +202323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -272001,83 +202547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -272302,83 +202771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -272603,83 +202995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -272905,83 +203220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -273207,83 +203445,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -273509,83 +203670,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -273811,83 +203895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -274112,83 +204119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -274413,83 +204343,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -274714,83 +204567,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -275015,83 +204791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -275316,83 +205015,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -275618,83 +205240,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -275919,83 +205464,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -276220,83 +205688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -276522,83 +205913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -276823,83 +206137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -277124,83 +206361,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -277425,83 +206585,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -277726,83 +206809,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -278027,83 +207033,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -278328,83 +207257,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -278630,83 +207482,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -278931,83 +207706,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -279232,83 +207930,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -279534,83 +208155,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -279835,83 +208379,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -280137,83 +208604,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -280438,83 +208828,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -280739,83 +209052,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -281040,83 +209276,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -281341,83 +209500,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -281643,83 +209725,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -281944,83 +209949,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -282245,83 +210173,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -282547,83 +210398,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -282848,83 +210622,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -283150,83 +210847,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -283452,83 +211072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -283753,83 +211296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -284054,83 +211520,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -284355,83 +211744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -284657,83 +211969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -284958,83 +212193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -285260,83 +212418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -285561,83 +212642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -285862,83 +212866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -286163,83 +213090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -286465,83 +213315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -286767,83 +213540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -287069,83 +213765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -287370,83 +213989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -287671,83 +214213,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -287972,83 +214437,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -288273,83 +214661,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -288574,83 +214885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -288875,83 +215109,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -289176,83 +215333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -289477,83 +215557,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -289778,83 +215781,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -290079,83 +216005,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -290380,83 +216229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -290681,83 +216453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -290982,83 +216677,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -291283,83 +216901,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -291584,83 +217125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -291885,83 +217349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -292186,83 +217573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -292487,83 +217797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -292788,83 +218021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -293089,83 +218245,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -293390,83 +218469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -293691,83 +218693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -293992,83 +218917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -294294,83 +219142,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -294596,83 +219367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -294897,83 +219591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -295198,83 +219815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -295499,83 +220039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -295800,83 +220263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -296102,83 +220488,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -296403,83 +220712,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -296704,83 +220936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -297005,83 +221160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -297307,83 +221385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -297609,83 +221610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -297910,83 +221834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -298211,83 +222058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -298512,83 +222282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -298813,83 +222506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -299114,83 +222730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -299415,83 +222954,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -299716,83 +223178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -300017,83 +223402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -300318,83 +223626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -300619,83 +223850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -300921,83 +224075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -301222,83 +224299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -301523,83 +224523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -301824,83 +224747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -302125,83 +224971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -302426,83 +225195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -302727,83 +225419,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -303028,83 +225643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -303329,83 +225867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -303630,83 +226091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -303931,83 +226315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -304232,83 +226539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -304533,83 +226763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: true - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_S_MX_B_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_S_MX_B_UserArgs.yaml index 0a44d8b8a13..cf2910b8852 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_S_MX_B_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bjlk_S_MX_B_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049, Device 0050] @@ -227,83 +227,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 10 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 1 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 1 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BBS_BH_Bias_BiasSrcD_GradB_AS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BBS_BH_Bias_BiasSrcD_GradB_AS_SAV_UserArgs.yaml index 4d5e1b24e5f..cf5b069475a 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BBS_BH_Bias_BiasSrcD_GradB_AS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BBS_BH_Bias_BiasSrcD_GradB_AS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -70,7 +70,7 @@ UseE: true UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 UseScaleCD: false - - 1LDSBuffer: 0 @@ -205,78 +205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -462,78 +390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -719,78 +575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -976,78 +760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs.yaml index d5d0b4e4a05..c38c3e20ccf 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -222,81 +222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -506,81 +431,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -790,81 +640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1074,81 +849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1358,81 +1058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1642,81 +1267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1926,81 +1476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2210,81 +1685,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2494,81 +1894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2778,81 +2103,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3062,81 +2312,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3346,81 +2521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3630,81 +2730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3914,81 +2939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4198,81 +3148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4482,81 +3357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4766,81 +3566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5050,81 +3775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5334,81 +3984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5618,81 +4193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5902,81 +4402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6186,81 +4611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6470,81 +4820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6754,81 +5029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7038,81 +5238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7322,81 +5447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7606,81 +5656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7890,81 +5865,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8174,81 +6074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8458,81 +6283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8742,81 +6492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9026,81 +6701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9310,81 +6910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9594,81 +7119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9878,81 +7328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10162,81 +7537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10446,81 +7746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10730,81 +7955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11014,81 +8164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11298,81 +8373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11582,81 +8582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11866,81 +8791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12150,81 +9000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12434,81 +9209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12718,81 +9418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13002,81 +9627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13286,81 +9836,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13570,81 +10045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13854,81 +10254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14138,81 +10463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14422,81 +10672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14706,81 +10881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14990,81 +11090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15274,81 +11299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15558,81 +11508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15842,81 +11717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16126,81 +11926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16410,81 +12135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16694,81 +12344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16978,81 +12553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17262,81 +12762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17546,81 +12971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17830,81 +13180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18114,81 +13389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18398,81 +13598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18682,81 +13807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18966,81 +14016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19250,81 +14225,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19534,81 +14434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19818,81 +14643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20102,81 +14852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20386,81 +15061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20670,81 +15270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20954,81 +15479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21238,81 +15688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21522,81 +15897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21806,81 +16106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22090,81 +16315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22374,81 +16524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22658,81 +16733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22942,81 +16942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23226,81 +17151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23518,83 +17368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23819,83 +17592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24120,83 +17816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24421,83 +18040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24722,83 +18264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25023,83 +18488,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25324,83 +18712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25625,83 +18936,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25926,83 +19160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26227,83 +19384,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26528,83 +19608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26829,83 +19832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27130,83 +20056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27431,83 +20280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27732,83 +20504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28034,83 +20729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28336,83 +20954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28637,83 +21178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28938,83 +21402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29239,83 +21626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29541,83 +21851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29843,83 +22076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30145,83 +22301,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30446,83 +22525,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30748,83 +22750,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31050,83 +22975,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31351,83 +23199,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31652,83 +23423,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31953,83 +23647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32254,83 +23871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32555,83 +24095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32857,83 +24320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33158,83 +24544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33459,83 +24768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33761,83 +24993,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34062,83 +25217,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34363,83 +25441,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34664,83 +25665,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34966,83 +25890,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35268,83 +26115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35569,83 +26339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35870,83 +26563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36171,83 +26787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36472,83 +27011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36773,83 +27235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37074,83 +27459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37375,83 +27683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37676,83 +27907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37977,83 +28131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38278,83 +28355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38579,83 +28579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38881,83 +28804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39183,83 +29029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39485,83 +29254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39786,83 +29478,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40087,83 +29702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40388,83 +29926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40690,83 +30151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40991,83 +30375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41292,83 +30599,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41593,83 +30823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41894,83 +31047,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42196,83 +31272,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42497,83 +31496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42798,83 +31720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43100,83 +31945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43402,83 +32170,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43703,83 +32394,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44004,83 +32618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44305,83 +32842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44606,83 +33066,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44908,83 +33291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45209,83 +33515,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45510,83 +33739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45811,83 +33963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46113,83 +34188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46414,83 +34412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46715,83 +34636,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47016,83 +34860,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47317,83 +35084,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47618,83 +35308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47919,83 +35532,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48220,83 +35756,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48522,83 +35981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48823,83 +36205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49124,83 +36429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49426,83 +36654,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49728,83 +36879,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50029,83 +37103,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50331,83 +37328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50632,83 +37552,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50933,83 +37776,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51234,83 +38000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51536,83 +38225,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51837,83 +38449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52138,83 +38673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52439,83 +38897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52740,83 +39121,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53041,83 +39345,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53342,83 +39569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53643,83 +39793,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53945,83 +40018,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54246,83 +40242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54547,83 +40466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54848,83 +40690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55150,83 +40915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55452,83 +41140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55754,83 +41365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56055,83 +41589,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56356,83 +41813,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56658,83 +42038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56960,83 +42263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57261,83 +42487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57562,83 +42711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57864,83 +42936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58165,83 +43160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58467,83 +43385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58768,83 +43609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59069,83 +43833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59370,83 +44057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59671,83 +44281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59972,83 +44505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60273,83 +44729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60574,83 +44953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60875,83 +45177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61176,83 +45401,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61478,83 +45626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61779,83 +45850,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62080,83 +46074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62381,83 +46298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62682,83 +46522,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62983,83 +46746,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63284,83 +46970,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63585,83 +47194,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63887,83 +47419,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64188,83 +47643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64489,83 +47867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64790,83 +48091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65091,83 +48315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65392,83 +48539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65694,83 +48764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65996,83 +48989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66297,83 +49213,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66598,83 +49437,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66899,83 +49661,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67200,83 +49885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67501,83 +50109,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67802,83 +50333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68103,83 +50557,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68404,83 +50781,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68705,83 +51005,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69006,83 +51229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69307,83 +51453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69608,83 +51677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69909,83 +51901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70210,83 +52125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70511,83 +52349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70812,83 +52573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71113,83 +52797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71415,83 +53022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71717,83 +53247,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72018,83 +53471,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72319,83 +53695,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72620,83 +53919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72921,83 +54143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73222,83 +54367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73523,83 +54591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73824,83 +54815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74125,83 +55039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74426,83 +55263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74727,83 +55487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75028,83 +55711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75329,83 +55935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75630,83 +56159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75931,83 +56383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76232,83 +56607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76533,83 +56831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76835,83 +57056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77136,83 +57280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77437,83 +57504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77738,83 +57728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78039,83 +57952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78340,83 +58176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78641,83 +58400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78942,83 +58624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79243,83 +58848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79544,83 +59072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79845,83 +59296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80146,83 +59520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80447,83 +59744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80748,83 +59968,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81049,83 +60192,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81350,83 +60416,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81651,83 +60640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81952,83 +60864,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82254,83 +61089,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82556,83 +61314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82858,83 +61539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83159,83 +61763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83460,83 +61987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83762,83 +62212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84064,83 +62437,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84365,83 +62661,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84667,83 +62886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84969,83 +63111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85270,83 +63335,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85571,83 +63559,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85872,83 +63783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86173,83 +64007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86474,83 +64231,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86776,83 +64456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87077,83 +64680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87378,83 +64904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87679,83 +65128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87980,83 +65352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88282,83 +65577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88583,83 +65801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88884,83 +66025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89185,83 +66249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89487,83 +66474,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89788,83 +66698,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90089,83 +66922,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90391,83 +67147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90692,83 +67371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90994,83 +67596,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91295,83 +67820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91596,83 +68044,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91897,83 +68268,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92199,83 +68493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92500,83 +68717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92802,83 +68942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93103,83 +69166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93404,83 +69390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93706,83 +69615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94008,83 +69840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94309,83 +70064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94610,83 +70288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94911,83 +70512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95212,83 +70736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95514,83 +70961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95816,83 +71186,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96117,83 +71410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96419,83 +71635,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96720,83 +71859,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97022,83 +72084,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97324,83 +72309,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97625,83 +72533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97926,83 +72757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98227,83 +72981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98529,83 +73206,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98830,83 +73430,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99131,83 +73654,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99432,83 +73878,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99733,83 +74102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100034,83 +74326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100336,83 +74551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100637,83 +74775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100939,83 +75000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101240,83 +75224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101541,83 +75448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101842,83 +75672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102144,83 +75897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102446,83 +76122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102748,83 +76347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103049,83 +76571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103350,83 +76795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103652,83 +77020,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103953,83 +77244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104255,83 +77469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104556,83 +77693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104857,83 +77917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105158,83 +78141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105459,83 +78365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105761,83 +78590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106063,83 +78815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106365,83 +79040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106667,83 +79265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106969,83 +79490,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107271,83 +79715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107573,83 +79940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107874,83 +80164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108175,83 +80388,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108476,83 +80612,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108777,83 +80836,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109078,83 +81060,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109380,83 +81285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109681,83 +81509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109982,83 +81733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110283,83 +81957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110585,83 +82182,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110887,83 +82407,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111188,83 +82631,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111490,83 +82856,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111791,83 +83080,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112092,83 +83304,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112394,83 +83529,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112696,83 +83754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112998,83 +83979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113299,83 +84203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113600,83 +84427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113901,83 +84651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114202,83 +84875,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114503,83 +85099,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114805,83 +85324,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115106,83 +85548,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115407,83 +85772,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115708,83 +85996,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116009,83 +86220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116310,83 +86444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116612,83 +86669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116914,83 +86894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117215,83 +87118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117517,83 +87343,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117819,83 +87568,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118121,83 +87793,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118422,83 +88017,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118723,83 +88241,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119025,83 +88466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119326,83 +88690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119628,83 +88915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119929,83 +89139,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120230,83 +89363,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120531,83 +89587,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120833,83 +89812,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121134,83 +90036,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121435,83 +90260,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121737,83 +90485,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122038,83 +90709,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122339,83 +90933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122640,83 +91157,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122941,83 +91381,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123242,83 +91605,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123543,83 +91829,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123844,83 +92053,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124145,83 +92277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124446,83 +92501,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124747,83 +92725,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125049,83 +92950,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125351,83 +93175,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125652,83 +93399,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125953,83 +93623,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126254,83 +93847,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126555,83 +94071,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126856,83 +94295,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127157,83 +94519,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127458,83 +94743,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127759,83 +94967,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128060,83 +95191,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128361,83 +95415,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128662,83 +95639,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128964,83 +95864,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129265,83 +96088,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129567,83 +96313,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129868,83 +96537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130169,83 +96761,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130470,83 +96985,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130771,83 +97209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131073,83 +97434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131374,83 +97658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131676,83 +97883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131977,83 +98107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132278,83 +98331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132579,83 +98555,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132880,83 +98779,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133181,83 +99003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133482,83 +99227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133783,83 +99451,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134084,83 +99675,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134385,83 +99899,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134686,83 +100123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134988,83 +100348,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135289,83 +100572,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135590,83 +100796,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135891,83 +101020,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136192,83 +101244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136494,83 +101469,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136796,83 +101694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137097,83 +101918,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137399,83 +102143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137700,83 +102367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138002,83 +102592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138303,83 +102816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138604,83 +103040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138905,83 +103264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139206,83 +103488,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139507,83 +103712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139809,83 +103937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140110,83 +104161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140412,83 +104386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140713,83 +104610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141014,83 +104834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141315,83 +105058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141617,83 +105283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141918,83 +105507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142219,83 +105731,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142521,83 +105956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142822,83 +106180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143124,83 +106405,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143425,83 +106629,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143726,83 +106853,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144028,83 +107078,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144329,83 +107302,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144630,83 +107526,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144931,83 +107750,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145233,83 +107975,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145534,83 +108199,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145835,83 +108423,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146136,83 +108647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146437,83 +108871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146738,83 +109095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147039,83 +109319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147341,83 +109544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147642,83 +109768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147943,83 +109992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148245,83 +110217,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148547,83 +110442,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148848,83 +110666,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149149,83 +110890,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149450,83 +111114,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149752,83 +111339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150053,83 +111563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150354,83 +111787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150656,83 +112012,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150958,83 +112237,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151259,83 +112461,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151560,83 +112685,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151861,83 +112909,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152162,83 +113133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152463,83 +113357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152764,83 +113581,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153065,83 +113805,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153366,83 +114029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153667,83 +114253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153968,83 +114477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154269,83 +114701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154570,83 +114925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154871,83 +115149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155172,83 +115373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155473,83 +115597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155775,83 +115822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156076,83 +116046,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156377,83 +116270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156678,83 +116494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156980,83 +116719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157282,83 +116944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157583,83 +117168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157885,83 +117393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158186,83 +117617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158487,83 +117841,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158789,83 +118066,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159090,83 +118290,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159391,83 +118514,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159693,83 +118739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159994,83 +118963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160296,83 +119188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160598,83 +119413,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160899,83 +119637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161200,83 +119861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161501,83 +120085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161803,83 +120310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162104,83 +120534,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162405,83 +120758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162706,83 +120982,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163007,83 +121206,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163308,83 +121430,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163609,83 +121654,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163910,83 +121878,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164211,83 +122102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164512,83 +122326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164813,83 +122550,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165114,83 +122774,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165415,83 +122998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165716,83 +123222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166018,83 +123447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166320,83 +123672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166621,83 +123896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166923,83 +124121,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167224,83 +124345,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167525,83 +124569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167826,83 +124793,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168127,83 +125017,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168428,83 +125241,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168729,83 +125465,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169030,83 +125689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169331,83 +125913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169632,83 +126137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169933,83 +126361,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170234,83 +126585,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170535,83 +126809,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170836,83 +127033,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171137,83 +127257,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171438,83 +127481,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171740,83 +127706,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172041,83 +127930,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172342,83 +128154,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172644,83 +128379,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172946,83 +128604,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173248,83 +128829,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173549,83 +129053,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173850,83 +129277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174152,83 +129502,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174454,83 +129727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174756,83 +129952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175057,83 +130176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175358,83 +130400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175660,83 +130625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175961,83 +130849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176262,83 +131073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176564,83 +131298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176865,83 +131522,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177166,83 +131746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177468,83 +131971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177769,83 +132195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178070,83 +132419,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178372,83 +132644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178673,83 +132868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178974,83 +133092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179275,83 +133316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179576,83 +133540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179877,83 +133764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180178,83 +133988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180479,83 +134212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180780,83 +134436,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181081,83 +134660,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181383,83 +134885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181685,83 +135110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181986,83 +135334,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182287,83 +135558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182588,83 +135782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182889,83 +136006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183190,83 +136230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183492,83 +136455,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183793,83 +136679,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184094,83 +136903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184395,83 +137127,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184697,83 +137352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184998,83 +137576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185299,83 +137800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185601,83 +138025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185902,83 +138249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186203,83 +138473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186504,83 +138697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186806,83 +138922,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187107,83 +139146,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187409,83 +139371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187710,83 +139595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188012,83 +139820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188313,83 +140044,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188614,83 +140268,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188915,83 +140492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189216,83 +140716,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189517,83 +140940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189819,83 +141165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190120,83 +141389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190422,83 +141614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190723,83 +141838,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191024,83 +142062,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191325,83 +142286,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191626,83 +142510,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191928,83 +142735,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192229,83 +142959,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192530,83 +143183,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192831,83 +143407,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193133,83 +143632,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193434,83 +143856,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193735,83 +144080,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194036,83 +144304,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194337,83 +144528,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194638,83 +144752,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194939,83 +144976,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195240,83 +145200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195541,83 +145424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195843,83 +145649,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196144,83 +145873,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196445,83 +146097,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196746,83 +146321,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197047,83 +146545,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197348,83 +146769,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197649,83 +146993,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197950,83 +147217,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198251,83 +147441,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198552,83 +147665,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198853,83 +147889,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199154,83 +148113,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199455,83 +148337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199756,83 +148561,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200057,83 +148785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200358,83 +149009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200659,83 +149233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200960,83 +149457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201261,83 +149681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201563,83 +149906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201865,83 +150131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202166,83 +150355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202467,83 +150579,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202768,83 +150803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203069,83 +151027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203371,83 +151252,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203672,83 +151476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203973,83 +151700,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204275,83 +151925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204577,83 +152150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204878,83 +152374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205180,83 +152599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205481,83 +152823,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205783,83 +153048,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206084,83 +153272,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206385,83 +153496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206686,83 +153720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206987,83 +153944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207288,83 +154168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207590,83 +154393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207891,83 +154617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208192,83 +154841,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208493,83 +155065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208794,83 +155289,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209095,83 +155513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209396,83 +155737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209697,83 +155961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209998,83 +156185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210299,83 +156409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210600,83 +156633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210901,83 +156857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211202,83 +157081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211504,83 +157306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211805,83 +157530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212107,83 +157755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212409,83 +157980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212710,83 +158204,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213011,83 +158428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213312,83 +158652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213613,83 +158876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213915,83 +159101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214216,83 +159325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214517,83 +159549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214818,83 +159773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215119,83 +159997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215420,83 +160221,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215722,83 +160446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216023,83 +160670,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216324,83 +160894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216625,83 +161118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216926,83 +161342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217227,83 +161566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217528,83 +161790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217829,83 +162014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218131,83 +162239,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218433,83 +162464,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218735,83 +162689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219037,83 +162914,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219338,83 +163138,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219639,83 +163362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219940,83 +163586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220242,83 +163811,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220543,83 +164035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220844,83 +164259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221145,83 +164483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221446,83 +164707,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221747,83 +164931,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222048,83 +165155,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222349,83 +165379,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222651,83 +165604,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222952,83 +165828,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223253,83 +166052,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223555,83 +166277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223856,83 +166501,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224157,83 +166725,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224458,83 +166949,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224759,83 +167173,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225060,83 +167397,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225361,83 +167621,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225662,83 +167845,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225963,83 +168069,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226264,83 +168293,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226565,83 +168517,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226867,83 +168742,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227169,83 +168967,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227470,83 +169191,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227772,83 +169416,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228074,83 +169641,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228376,83 +169866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228677,83 +170090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228978,83 +170314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229279,83 +170538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229580,83 +170762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229881,83 +170986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230182,83 +171210,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230484,83 +171435,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230785,83 +171659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231087,83 +171884,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231388,83 +172108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231690,83 +172333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231992,83 +172558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232293,83 +172782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232595,83 +173007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232897,83 +173232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233198,83 +173456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233499,83 +173680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233801,83 +173905,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234103,83 +174130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234404,83 +174354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234705,83 +174578,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235006,83 +174802,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235307,83 +175026,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235609,83 +175251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235911,83 +175476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236212,83 +175700,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236513,83 +175924,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236814,83 +176148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237115,83 +176372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237416,83 +176596,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237717,83 +176820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238018,83 +177044,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238319,83 +177268,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238620,83 +177492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238921,83 +177716,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239223,83 +177941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239524,83 +178165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239825,83 +178389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240126,83 +178613,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240427,83 +178837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240728,83 +179061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241029,83 +179285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241330,83 +179509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241631,83 +179733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241932,83 +179957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242233,83 +180181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242535,83 +180406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242837,83 +180631,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243138,83 +180855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243440,83 +181080,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243742,83 +181305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244043,83 +181529,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244344,83 +181753,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244645,83 +181977,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244946,83 +182201,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245247,83 +182425,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245548,83 +182649,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245850,83 +182874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246151,83 +183098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246453,83 +183323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246754,83 +183547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247055,83 +183771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247356,83 +183995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247658,83 +184220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247959,83 +184444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248260,83 +184668,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248561,83 +184892,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248863,83 +185117,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249165,83 +185342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249466,83 +185566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BSS_BH_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BSS_BH_Bias_HAS_SAV_UserArgs.yaml index dc65ec3180f..aefd1e60e65 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BSS_BH_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_BSS_BH_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -834,83 +680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1135,83 +904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1436,83 +1128,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1737,83 +1352,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2038,83 +1576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2339,83 +1800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2640,83 +2024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2941,83 +2248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3243,83 +2473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3544,83 +2697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3845,83 +2921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4146,83 +3145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4447,83 +3369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4749,83 +3594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5050,83 +3818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5351,83 +4042,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5652,83 +4266,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5953,83 +4490,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6254,83 +4714,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6556,83 +4939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6857,83 +5163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7158,83 +5387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7459,83 +5611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7760,83 +5835,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8061,83 +6059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8363,83 +6284,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8664,83 +6508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8966,83 +6733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9267,83 +6957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9568,83 +7181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9869,83 +7405,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10170,83 +7629,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10472,83 +7854,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10773,83 +8078,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11074,83 +8302,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11376,83 +8527,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11677,83 +8751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11978,83 +8975,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12279,83 +9199,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12580,83 +9423,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12881,83 +9647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13182,83 +9871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13483,83 +10095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13784,83 +10319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14085,83 +10543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14386,83 +10767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14687,83 +10991,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14988,83 +11215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15289,83 +11439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15590,83 +11663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15891,83 +11887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16192,83 +12111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16493,83 +12335,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16794,83 +12559,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17095,83 +12783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17396,83 +13007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17698,83 +13232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18000,83 +13457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18302,83 +13682,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18603,83 +13906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18904,83 +14130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19205,83 +14354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19507,83 +14579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19808,83 +14803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20109,83 +15027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20410,83 +15251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20711,83 +15475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21013,83 +15700,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21315,83 +15925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21616,83 +16149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21917,83 +16373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22218,83 +16597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22519,83 +16821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22820,83 +17045,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23122,83 +17270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23423,83 +17494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23725,83 +17719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24026,83 +17943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24327,83 +18167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24628,83 +18391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24929,83 +18615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25231,83 +18840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25532,83 +19064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25833,83 +19288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26135,83 +19513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26436,83 +19737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26737,83 +19961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27039,83 +20186,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27340,83 +20410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27641,83 +20634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27942,83 +20858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28243,83 +21082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28544,83 +21306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28845,83 +21530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29147,83 +21755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29448,83 +21979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29749,83 +22203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30050,83 +22427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30352,83 +22652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30654,83 +22877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30956,83 +23102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31257,83 +23326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31558,83 +23550,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31860,83 +23775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32162,83 +24000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32464,83 +24225,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32765,83 +24449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33066,83 +24673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33367,83 +24897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33669,83 +25122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33970,83 +25346,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34271,83 +25570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34572,83 +25794,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34873,83 +26018,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35174,83 +26242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35475,83 +26466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35776,83 +26690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36077,83 +26914,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36378,83 +27138,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36679,83 +27362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36981,83 +27587,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37283,83 +27812,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37584,83 +28036,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37886,83 +28261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38187,83 +28485,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38489,83 +28710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38791,83 +28935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39092,83 +29159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39394,83 +29384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39695,83 +29608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39996,83 +29832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40297,83 +30056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40599,83 +30281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40900,83 +30505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41201,83 +30729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41502,83 +30953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41803,83 +31177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42104,83 +31401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42405,83 +31625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42707,83 +31850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43009,83 +32075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43311,83 +32300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43612,83 +32524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43913,83 +32748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44214,83 +32972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44515,83 +33196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44816,83 +33420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45117,83 +33644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45418,83 +33868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45720,83 +34093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46021,83 +34317,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46322,83 +34541,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46623,83 +34765,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46924,83 +34989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47226,83 +35214,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47527,83 +35438,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47828,83 +35662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48129,83 +35886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48430,83 +36110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48731,83 +36334,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49032,83 +36558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49333,83 +36782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49635,83 +37007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49936,83 +37231,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50238,83 +37456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50539,83 +37680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50840,83 +37904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51141,83 +38128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51442,83 +38352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51743,83 +38576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52044,83 +38800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52345,83 +39024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52646,83 +39248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52947,83 +39472,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53248,83 +39696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53549,83 +39920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53850,83 +40144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54151,83 +40368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54452,83 +40592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54753,83 +40816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55054,83 +41040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55356,83 +41265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55658,83 +41490,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55960,83 +41715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56261,83 +41939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56562,83 +42163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56863,83 +42387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57164,83 +42611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57465,83 +42835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57766,83 +43059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58067,83 +43283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58368,83 +43507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58670,83 +43732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58971,83 +43956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59273,83 +44181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59574,83 +44405,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59875,83 +44629,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60177,83 +44854,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60478,83 +45078,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60779,83 +45302,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61080,83 +45526,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61381,83 +45750,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61682,83 +45974,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61983,83 +46198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62284,83 +46422,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62585,83 +46646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62886,83 +46870,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63188,83 +47095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63490,83 +47320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63791,83 +47544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64092,83 +47768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64393,83 +47992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64694,83 +48216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64995,83 +48440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65296,83 +48664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65597,83 +48888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65898,83 +49112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66199,83 +49336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66500,83 +49560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66801,83 +49784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67102,83 +50008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67403,83 +50232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67704,83 +50456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68005,83 +50680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68307,83 +50905,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68609,83 +51130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68911,83 +51355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69212,83 +51579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69514,83 +51804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69815,83 +52028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70116,83 +52252,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70417,83 +52476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70719,83 +52701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71020,83 +52925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71321,83 +53149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71622,83 +53373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71923,83 +53597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72224,83 +53821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72525,83 +54045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72826,83 +54269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73127,83 +54493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73428,83 +54717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73729,83 +54941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74030,83 +55165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74332,83 +55390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74633,83 +55614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74935,83 +55839,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75236,83 +56063,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75538,83 +56288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75840,83 +56513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76142,83 +56738,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76443,83 +56962,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76745,83 +57187,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77047,83 +57412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77349,83 +57637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77650,83 +57861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77952,83 +58086,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78253,83 +58310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78554,83 +58534,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78856,83 +58759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79157,83 +58983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79458,83 +59207,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79759,83 +59431,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80060,83 +59655,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80361,83 +59879,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80662,83 +60103,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80963,83 +60327,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81264,83 +60551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81565,83 +60775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81866,83 +60999,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82167,83 +61223,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82468,83 +61447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82769,83 +61671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83070,83 +61895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83372,83 +62120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83674,83 +62345,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83975,83 +62569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84276,83 +62793,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84577,83 +63017,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84878,83 +63241,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85179,83 +63465,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85480,83 +63689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85782,83 +63914,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86083,83 +64138,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86384,83 +64362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86685,83 +64586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86987,83 +64811,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87288,83 +65035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87589,83 +65259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87890,83 +65483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88191,83 +65707,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88493,83 +65932,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88794,83 +66156,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89095,83 +66380,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89397,83 +66605,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89698,83 +66829,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90000,83 +67054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90301,83 +67278,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90602,83 +67502,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90904,83 +67727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91205,83 +67951,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91506,83 +68175,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91807,83 +68399,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92108,83 +68623,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92409,83 +68847,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92711,83 +69072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93012,83 +69296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93313,83 +69520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93614,83 +69744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93915,83 +69968,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94216,83 +70192,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94518,83 +70417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94819,83 +70641,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95121,83 +70866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95422,83 +71090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95724,83 +71315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96026,83 +71540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96328,83 +71765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96630,83 +71990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96931,83 +72214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97233,83 +72439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97534,83 +72663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97835,83 +72887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98136,83 +73111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98437,83 +73335,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98738,83 +73559,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99039,83 +73783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99340,83 +74007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99641,83 +74231,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99942,83 +74455,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100243,83 +74679,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100544,83 +74903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100846,83 +75128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101148,83 +75353,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101449,83 +75577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101750,83 +75801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102051,83 +76025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102352,83 +76249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102653,83 +76473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102955,83 +76698,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103256,83 +76922,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103557,83 +77146,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103858,83 +77370,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104159,83 +77594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104461,83 +77819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104762,83 +78043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105063,83 +78267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105365,83 +78492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105666,83 +78716,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105968,83 +78941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106269,83 +79165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106570,83 +79389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106871,83 +79613,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107172,83 +79837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107473,83 +80061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107774,83 +80285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108075,83 +80509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108377,83 +80734,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108678,83 +80958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108979,83 +81182,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109280,83 +81406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109582,83 +81631,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109883,83 +81855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110184,83 +82079,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110485,83 +82303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110786,83 +82527,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111087,83 +82751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111389,83 +82976,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111690,83 +83200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111991,83 +83424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112293,83 +83649,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112594,83 +83873,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112896,83 +84098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113198,83 +84323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113499,83 +84547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113801,83 +84772,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114102,83 +84996,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114404,83 +85221,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114705,83 +85445,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115006,83 +85669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115307,83 +85893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115608,83 +86117,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115909,83 +86341,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116210,83 +86565,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116511,83 +86789,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116812,83 +87013,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117113,83 +87237,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117414,83 +87461,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117715,83 +87685,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118016,83 +87909,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118317,83 +88133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118619,83 +88358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118920,83 +88582,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119221,83 +88806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119523,83 +89031,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119824,83 +89255,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120125,83 +89479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120426,83 +89703,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120727,83 +89927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121028,83 +90151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121329,83 +90375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121630,83 +90599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121932,83 +90824,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122233,83 +91048,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122534,83 +91272,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122835,83 +91496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123136,83 +91720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123437,83 +91944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123738,83 +92168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124040,83 +92393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124341,83 +92617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124642,83 +92841,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124943,83 +93065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125244,83 +93289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125546,83 +93514,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125847,83 +93738,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126148,83 +93962,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126449,83 +94186,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126751,83 +94411,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127052,83 +94635,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127353,83 +94859,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127655,83 +95084,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127956,83 +95308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128258,83 +95533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128559,83 +95757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128860,83 +95981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129161,83 +96205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129462,83 +96429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129763,83 +96653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130064,83 +96877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130366,83 +97102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130667,83 +97326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130968,83 +97550,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131270,83 +97775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131571,83 +97999,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131872,83 +98223,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132173,83 +98447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132474,83 +98671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132775,83 +98895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133077,83 +99120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133378,83 +99344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133679,83 +99568,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133980,83 +99792,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134282,83 +100017,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134583,83 +100241,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134884,83 +100465,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135185,83 +100689,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135486,83 +100913,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135787,83 +101137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136088,83 +101361,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136390,83 +101586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136691,83 +101810,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136993,83 +102035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137294,83 +102259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137596,83 +102484,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137897,83 +102708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138199,83 +102933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138500,83 +103157,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138802,83 +103382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139103,83 +103606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139404,83 +103830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139705,83 +104054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140006,83 +104278,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140307,83 +104502,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140608,83 +104726,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140909,83 +104950,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141210,83 +105174,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141511,83 +105398,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141812,83 +105622,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142113,83 +105846,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142414,83 +106070,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142716,83 +106295,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143017,83 +106519,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143318,83 +106743,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143620,83 +106968,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143921,83 +107192,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144222,83 +107416,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144523,83 +107640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144825,83 +107865,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145126,83 +108089,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145427,83 +108313,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145729,83 +108538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146031,83 +108763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146332,83 +108987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146634,83 +109212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146935,83 +109436,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147236,83 +109660,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147538,83 +109885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147840,83 +110110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148141,83 +110334,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148443,83 +110559,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148744,83 +110783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149045,83 +111007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149347,83 +111232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149648,83 +111456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149950,83 +111681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150252,83 +111906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150553,83 +112130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150854,83 +112354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151156,83 +112579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151457,83 +112803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151758,83 +113027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152059,83 +113251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152360,83 +113475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152662,83 +113700,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152963,83 +113924,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153265,83 +114149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153567,83 +114374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153868,83 +114598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154169,83 +114822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154471,83 +115047,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154773,83 +115272,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155074,83 +115496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155375,83 +115720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155676,83 +115944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155977,83 +116168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156278,83 +116392,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156579,83 +116616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156880,83 +116840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157181,83 +117064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157482,83 +117288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157783,83 +117512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158084,83 +117736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158386,83 +117961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158687,83 +118185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158989,83 +118410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159290,83 +118634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159591,83 +118858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159892,83 +119082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160194,83 +119307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160495,83 +119531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160796,83 +119755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161097,83 +119979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161399,83 +120204,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161700,83 +120428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162001,83 +120652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162302,83 +120876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162603,83 +121100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162904,83 +121324,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163206,83 +121549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163508,83 +121774,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163809,83 +121998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164110,83 +122222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164411,83 +122446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164713,83 +122671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165014,83 +122895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165315,83 +123119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165616,83 +123343,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165917,83 +123567,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166219,83 +123792,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166521,83 +124017,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166822,83 +124241,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167124,83 +124466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167426,83 +124691,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167727,83 +124915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168028,83 +125139,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168329,83 +125363,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168630,83 +125587,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168932,83 +125812,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169233,83 +126036,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169534,83 +126260,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169836,83 +126485,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170137,83 +126709,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170439,83 +126934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170740,83 +127158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171041,83 +127382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171342,83 +127606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171644,83 +127831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171945,83 +128055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172246,83 +128279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172548,83 +128504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172849,83 +128728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173150,83 +128952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173451,83 +129176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173753,83 +129401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174054,83 +129625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174355,83 +129849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174656,83 +130073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174957,83 +130297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175259,83 +130522,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175560,83 +130746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175861,83 +130970,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176162,83 +131194,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176463,83 +131418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176764,83 +131642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177065,83 +131866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177367,83 +132091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177669,83 +132316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177971,83 +132541,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178272,83 +132765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178573,83 +132989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178875,83 +133214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179176,83 +133438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179478,83 +133663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179779,83 +133887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180081,83 +134112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180382,83 +134336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180683,83 +134560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180984,83 +134784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181285,83 +135008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181586,83 +135232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181888,83 +135457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182189,83 +135681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182490,83 +135905,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182792,83 +136130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183093,83 +136354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183394,83 +136578,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183696,83 +136803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183997,83 +137027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184298,83 +137251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184599,83 +137475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184900,83 +137699,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185201,83 +137923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185502,83 +138147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185803,83 +138371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186104,83 +138595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186405,83 +138819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186707,83 +139044,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187008,83 +139268,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187310,83 +139493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187611,83 +139717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187912,83 +139941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188213,83 +140165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188514,83 +140389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188815,83 +140613,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189116,83 +140837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189417,83 +141061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189718,83 +141285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190019,83 +141509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190320,83 +141733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190621,83 +141957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190922,83 +142181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191224,83 +142406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191525,83 +142630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191826,83 +142854,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192127,83 +143078,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192428,83 +143302,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192729,83 +143526,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193030,83 +143750,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193331,83 +143974,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193632,83 +144198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193934,83 +144423,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194235,83 +144647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194536,83 +144871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194838,83 +145096,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195140,83 +145321,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195442,83 +145546,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195743,83 +145770,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196044,83 +145994,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196345,83 +146218,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196646,83 +146442,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196947,83 +146666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197248,83 +146890,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197550,83 +147115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197851,83 +147339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198152,83 +147563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198454,83 +147788,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198755,83 +148012,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199056,83 +148236,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199357,83 +148460,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199659,83 +148685,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199960,83 +148909,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200261,83 +149133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200562,83 +149357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200863,83 +149581,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201165,83 +149806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201466,83 +150030,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201768,83 +150255,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202069,83 +150479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202371,83 +150704,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202672,83 +150928,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202974,83 +151153,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203275,83 +151377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203576,83 +151601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203877,83 +151825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204178,83 +152049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204479,83 +152273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204781,83 +152498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205082,83 +152722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205383,83 +152946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205684,83 +153170,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205985,83 +153394,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206287,83 +153619,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206588,83 +153843,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206889,83 +154067,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207190,83 +154291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207492,83 +154516,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207794,83 +154741,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208095,83 +154965,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208397,83 +155190,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208698,83 +155414,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209000,83 +155639,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209301,83 +155863,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209602,83 +156087,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209904,83 +156312,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210206,83 +156537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210507,83 +156761,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210808,83 +156985,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211109,83 +157209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211411,83 +157434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211712,83 +157658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212014,83 +157883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212315,83 +158107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212616,83 +158331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212917,83 +158555,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213219,83 +158780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213521,83 +159005,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213822,83 +159229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214123,83 +159453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214424,83 +159677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214725,83 +159901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215027,83 +160126,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215329,83 +160351,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215630,83 +160575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215931,83 +160799,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216232,83 +161023,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216534,83 +161248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216835,83 +161472,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217136,83 +161696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217437,83 +161920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217739,83 +162145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218041,83 +162370,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218343,83 +162595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218644,83 +162819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218945,83 +163043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219246,83 +163267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219548,83 +163492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219849,83 +163716,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220150,83 +163940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220452,83 +164165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220753,83 +164389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221054,83 +164613,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221355,83 +164837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221657,83 +165062,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221958,83 +165286,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222260,83 +165511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222562,83 +165736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222863,83 +165960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223164,83 +166184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223465,83 +166408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223767,83 +166633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224069,83 +166858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224371,83 +167083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224673,83 +167308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224974,83 +167532,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225275,83 +167756,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225577,83 +167981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225878,83 +168205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226179,83 +168429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226480,83 +168653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226781,83 +168877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227083,83 +169102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227384,83 +169326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227685,83 +169550,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227987,83 +169775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228288,83 +169999,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228590,83 +170224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228891,83 +170448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229192,83 +170672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229494,83 +170897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229795,83 +171121,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230097,83 +171346,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230399,83 +171571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230700,83 +171795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231002,83 +172020,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231303,83 +172244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231604,83 +172468,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231905,83 +172692,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232207,83 +172917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232508,83 +173141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232809,83 +173365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233110,83 +173589,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233412,83 +173814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233713,83 +174038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234015,83 +174263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234316,83 +174487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234617,83 +174711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234919,83 +174936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235220,83 +175160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235521,83 +175384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235822,83 +175608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236123,83 +175832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236425,83 +176057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236727,83 +176282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237028,83 +176506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237329,83 +176730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237631,83 +176955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237933,83 +177180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238235,83 +177405,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238536,83 +177629,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238837,83 +177853,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239138,83 +178077,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239439,83 +178301,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239740,83 +178525,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240041,83 +178749,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240343,83 +178974,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240644,83 +179198,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240945,83 +179422,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241246,83 +179646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241548,83 +179871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241850,83 +180096,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242151,83 +180320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242453,83 +180545,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242755,83 +180770,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243056,83 +180994,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243357,83 +181218,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243658,83 +181442,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243959,83 +181666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244261,83 +181891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244563,83 +182116,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244865,83 +182341,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245167,83 +182566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245468,83 +182790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245770,83 +183015,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246071,83 +183239,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246373,83 +183464,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246674,83 +183688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246975,83 +183912,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247277,83 +184137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247579,83 +184362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247880,83 +184586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248182,83 +184811,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248483,83 +185035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248784,83 +185259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249085,83 +185483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 7 - DataTypeA: 7 - DataTypeAmaxD: 0 - DataTypeB: 7 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_DB_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_DB_UserArgs.yaml index d7352f5f634..498ec4eee00 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_DB_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_DB_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -67,7 +67,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 0 - - 1LDSBuffer: 0 ActivationAlt: false @@ -194,74 +194,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 1 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 1 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - DataType: 1 - DestDataType: 1 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 0 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -440,74 +372,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 1 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 1 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - DataType: 1 - DestDataType: 1 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 0 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8B8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8B8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml index 900ad6ab24d..9c21cbadf67 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8B8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8B8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -834,83 +680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1135,83 +904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1436,83 +1128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1737,83 +1352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2038,83 +1576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2339,83 +1800,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2640,83 +2024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2941,83 +2248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3242,83 +2472,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3543,83 +2696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3844,83 +2920,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4145,83 +3144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4446,83 +3368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4747,83 +3592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5048,83 +3816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5349,83 +4040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5650,83 +4264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5951,83 +4488,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6252,83 +4712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6553,83 +4936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6854,83 +5160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7155,83 +5384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7456,83 +5608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7757,83 +5832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8058,83 +6056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8359,83 +6280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8661,83 +6505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8963,83 +6730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9264,83 +6954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9565,83 +7178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9866,83 +7402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10167,83 +7626,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10468,83 +7850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10769,83 +8074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11070,83 +8298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11371,83 +8522,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11672,83 +8746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11973,83 +8970,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12274,83 +9194,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12575,83 +9418,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12876,83 +9642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13177,83 +9866,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13479,83 +10091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13780,83 +10315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14081,83 +10539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14382,83 +10763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14684,83 +10988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14986,83 +11213,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15287,83 +11437,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15588,83 +11661,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15889,83 +11885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16190,83 +12109,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16491,83 +12333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16793,83 +12558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17094,83 +12782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17395,83 +13006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17696,83 +13230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17997,83 +13454,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18298,83 +13678,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18600,83 +13903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18901,83 +14127,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19203,83 +14352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19504,83 +14576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19806,83 +14801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20107,83 +15025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20408,83 +15249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20709,83 +15473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21010,83 +15697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21311,83 +15921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21612,83 +16145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21913,83 +16369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22215,83 +16594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22516,83 +16818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22818,83 +17043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23119,83 +17267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23420,83 +17491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23721,83 +17715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24022,83 +17939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24323,83 +18163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24624,83 +18387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24925,83 +18611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25226,83 +18835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25527,83 +19059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25828,83 +19283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26129,83 +19507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26430,83 +19731,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26732,83 +19956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27034,83 +20181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27336,83 +20406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27637,83 +20630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27938,83 +20854,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28239,83 +21078,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28540,83 +21302,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28841,83 +21526,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29143,83 +21751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29444,83 +21975,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29746,83 +22200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30047,83 +22424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30349,83 +22649,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30651,83 +22874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30952,83 +23098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31253,83 +23322,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31554,83 +23546,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31855,83 +23770,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32156,83 +23994,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32457,83 +24218,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32758,83 +24442,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33059,83 +24666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33361,83 +24891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33663,83 +25116,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33964,83 +25340,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34265,83 +25564,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34566,83 +25788,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34868,83 +26013,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35170,83 +26238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35471,83 +26462,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35772,83 +26686,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36073,83 +26910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36375,83 +27135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36676,83 +27359,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36977,83 +27583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37278,83 +27807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37580,83 +28032,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37882,83 +28257,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38184,83 +28482,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38485,83 +28706,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38786,83 +28930,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39087,83 +29154,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39388,83 +29378,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39689,83 +29602,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39990,83 +29826,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40291,83 +30050,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40592,83 +30274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40893,83 +30498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41194,83 +30722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41495,83 +30946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41796,83 +31170,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42097,83 +31394,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42398,83 +31618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42699,83 +31842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43000,83 +32066,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43302,83 +32291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43603,83 +32515,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43904,83 +32739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44205,83 +32963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44506,83 +33187,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44808,83 +33412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45110,83 +33637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45411,83 +33861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45712,83 +34085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46014,83 +34310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46315,83 +34534,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46617,83 +34759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46919,83 +34984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47220,83 +35208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47521,83 +35432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47822,83 +35656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48123,83 +35880,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48425,83 +36105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48727,83 +36330,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49028,83 +36554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49329,83 +36778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49630,83 +37002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49931,83 +37226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50232,83 +37450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50533,83 +37674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50835,83 +37899,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51136,83 +38123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51437,83 +38347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51738,83 +38571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52039,83 +38795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52341,83 +39020,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52642,83 +39244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52943,83 +39468,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53244,83 +39692,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53546,83 +39917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53847,83 +40141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54148,83 +40365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54449,83 +40589,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54750,83 +40813,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55051,83 +41037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55352,83 +41261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55654,83 +41486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55955,83 +41710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56256,83 +41934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56557,83 +42158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56858,83 +42382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57160,83 +42607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57462,83 +42832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57763,83 +43056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58064,83 +43280,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58365,83 +43504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58666,83 +43728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58967,83 +43952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59268,83 +44176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59569,83 +44400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59870,83 +44624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60171,83 +44848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60472,83 +45072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60773,83 +45296,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61074,83 +45520,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61375,83 +45744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61676,83 +45968,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61977,83 +46192,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62278,83 +46416,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62579,83 +46640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62880,83 +46864,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63181,83 +47088,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63483,83 +47313,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63784,83 +47537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64085,83 +47761,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64386,83 +47985,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64687,83 +48209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64988,83 +48433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65289,83 +48657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65591,83 +48882,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65892,83 +49106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66193,83 +49330,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66494,83 +49554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66796,83 +49779,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67097,83 +50003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67398,83 +50227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67699,83 +50451,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68000,83 +50675,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68301,83 +50899,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68602,83 +51123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68904,83 +51348,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69206,83 +51573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69507,83 +51797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69809,83 +52022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70110,83 +52246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70411,83 +52470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70712,83 +52694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71014,83 +52919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71315,83 +53143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71617,83 +53368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71919,83 +53593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72220,83 +53817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72521,83 +54041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72822,83 +54265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73123,83 +54489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73424,83 +54713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73725,83 +54937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74026,83 +55161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74327,83 +55385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74628,83 +55609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74929,83 +55833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75230,83 +56057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75532,83 +56282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75834,83 +56507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76136,83 +56732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76437,83 +56956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76738,83 +57180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77039,83 +57404,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77340,83 +57628,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77641,83 +57852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77942,83 +58076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78243,83 +58300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78544,83 +58524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78845,83 +58748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79146,83 +58972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79447,83 +59196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79749,83 +59421,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80050,83 +59645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80351,83 +59869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80653,83 +60094,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80954,83 +60318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81255,83 +60542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81557,83 +60767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81859,83 +60992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82160,83 +61216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82461,83 +61440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82762,83 +61664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83063,83 +61888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83364,83 +62112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83666,83 +62337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83967,83 +62561,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84268,83 +62785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84569,83 +63009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84870,83 +63233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85171,83 +63457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85473,83 +63682,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85774,83 +63906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86075,83 +64130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86377,83 +64355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86679,83 +64580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86980,83 +64804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87281,83 +65028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87583,83 +65253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87885,83 +65478,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88186,83 +65702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88488,83 +65927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88789,83 +66151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89090,83 +66375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89391,83 +66599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89693,83 +66824,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89995,83 +67049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90296,83 +67273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90597,83 +67497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90898,83 +67721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91199,83 +67945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91500,83 +68169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91801,83 +68393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92102,83 +68617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92403,83 +68841,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92704,83 +69065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93005,83 +69289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93306,83 +69513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93607,83 +69737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93908,83 +69961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94209,83 +70185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94510,83 +70409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94811,83 +70633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95112,83 +70857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95413,83 +71081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95715,83 +71306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96016,83 +71530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96317,83 +71754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96618,83 +71978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96919,83 +72202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97220,83 +72426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97521,83 +72650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97822,83 +72874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98123,83 +73098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98425,83 +73323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98726,83 +73547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99027,83 +73771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99328,83 +73995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99629,83 +74219,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99930,83 +74443,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100231,83 +74667,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100532,83 +74891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100833,83 +75115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101134,83 +75339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101435,83 +75563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101736,83 +75787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102037,83 +76011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102338,83 +76235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102639,83 +76459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102940,83 +76683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103241,83 +76907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103542,83 +77131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103843,83 +77355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104145,83 +77580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104446,83 +77804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104748,83 +78029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105049,83 +78253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105350,83 +78477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105651,83 +78701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105952,83 +78925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106254,83 +79150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106555,83 +79374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106857,83 +79599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107158,83 +79823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107460,83 +80048,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107762,83 +80273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108064,83 +80498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108366,83 +80723,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108667,83 +80947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108968,83 +81171,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109269,83 +81395,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109570,83 +81619,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109871,83 +81843,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110172,83 +82067,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110473,83 +82291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110774,83 +82515,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111075,83 +82739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111376,83 +82963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111678,83 +83188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111979,83 +83412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112280,83 +83636,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112582,83 +83861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112883,83 +84085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113184,83 +84309,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113485,83 +84533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113786,83 +84757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114087,83 +84981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114388,83 +85205,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114689,83 +85429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114990,83 +85653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115291,83 +85877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115592,83 +86101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115893,83 +86325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116194,83 +86549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116496,83 +86774,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116797,83 +86998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117098,83 +87222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117399,83 +87446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117701,83 +87671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118003,83 +87896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118304,83 +88120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118605,83 +88344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118907,83 +88569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119209,83 +88794,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119510,83 +89018,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119811,83 +89242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120112,83 +89466,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120413,83 +89690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120715,83 +89915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121017,83 +90140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121318,83 +90364,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121619,83 +90588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121921,83 +90813,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122223,83 +91038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122524,83 +91262,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122825,83 +91486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123126,83 +91710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123427,83 +91934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123728,83 +92158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124029,83 +92382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124331,83 +92607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124633,83 +92832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124935,83 +93057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125236,83 +93281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125537,83 +93505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125838,83 +93729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126139,83 +93953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126440,83 +94177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126741,83 +94401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127042,83 +94625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127343,83 +94849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127644,83 +95073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127945,83 +95297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128246,83 +95521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128547,83 +95745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128848,83 +95969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129149,83 +96193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129450,83 +96417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129751,83 +96641,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130052,83 +96865,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130354,83 +97090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130655,83 +97314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130956,83 +97538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131257,83 +97762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131558,83 +97986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131859,83 +98210,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132160,83 +98434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132461,83 +98658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132762,83 +98882,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133064,83 +99107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133366,83 +99332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133667,83 +99556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133968,83 +99780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134269,83 +100004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134570,83 +100228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134871,83 +100452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135172,83 +100676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135473,83 +100900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135774,83 +101124,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136075,83 +101348,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136377,83 +101573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136679,83 +101798,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136980,83 +102022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137281,83 +102246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137582,83 +102470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137883,83 +102694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138184,83 +102918,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138485,83 +103142,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138787,83 +103367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139088,83 +103591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139389,83 +103815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139690,83 +104039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139991,83 +104263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140292,83 +104487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140593,83 +104711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140894,83 +104935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141195,83 +105159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141496,83 +105383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141797,83 +105607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142098,83 +105831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142399,83 +106055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142700,83 +106279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143001,83 +106503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143303,83 +106728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143604,83 +106952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143906,83 +107177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144207,83 +107401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144509,83 +107626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144811,83 +107851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145112,83 +108075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145413,83 +108299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145715,83 +108524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146016,83 +108748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146317,83 +108972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146618,83 +109196,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146919,83 +109420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147220,83 +109644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147521,83 +109868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147822,83 +110092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148123,83 +110316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148424,83 +110540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148726,83 +110765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149028,83 +110990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149329,83 +111214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149631,83 +111439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149932,83 +111663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150233,83 +111887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150534,83 +112111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150836,83 +112336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151137,83 +112560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151438,83 +112784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151740,83 +113009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152041,83 +113233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152342,83 +113457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152643,83 +113681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152944,83 +113905,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153245,83 +114129,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153546,83 +114353,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153847,83 +114577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154148,83 +114801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154450,83 +115026,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154752,83 +115251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155053,83 +115475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155355,83 +115700,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155656,83 +115924,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155957,83 +116148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156258,83 +116372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156560,83 +116597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156861,83 +116821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157162,83 +117045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157463,83 +117269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157765,83 +117494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158066,83 +117718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158368,83 +117943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158669,83 +118167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158970,83 +118391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159272,83 +118616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159573,83 +118840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159874,83 +119064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160175,83 +119288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160476,83 +119512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160777,83 +119736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161078,83 +119960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161379,83 +120184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161680,83 +120408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161981,83 +120632,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162283,83 +120857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162584,83 +121081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162885,83 +121305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163187,83 +121530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163488,83 +121754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163789,83 +121978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164090,83 +122202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164391,83 +122426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164692,83 +122650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164994,83 +122875,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165295,83 +123099,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165596,83 +123323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165897,83 +123547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166198,83 +123771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166499,83 +123995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166800,83 +124219,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167102,83 +124444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167404,83 +124669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167705,83 +124893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168006,83 +125117,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168307,83 +125341,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168608,83 +125565,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168909,83 +125789,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169210,83 +126013,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169512,83 +126238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169813,83 +126462,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170114,83 +126686,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170415,83 +126910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170716,83 +127134,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171017,83 +127358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171318,83 +127582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171619,83 +127806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171920,83 +128030,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172221,83 +128254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172523,83 +128479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172824,83 +128703,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173125,83 +128927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173427,83 +129152,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173729,83 +129377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174030,83 +129601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174331,83 +129825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174632,83 +130049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174934,83 +130274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 17 - DataTypeA: 17 - DataTypeAmaxD: 0 - DataTypeB: 17 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml index 5d0e9a6d1c3..90e22ed5761 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8BS_BH_BiasSB_HAS_SAB_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -834,83 +680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1135,83 +904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1436,83 +1128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1737,83 +1352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2038,83 +1576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2339,83 +1800,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2640,83 +2024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2941,83 +2248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3242,83 +2472,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3543,83 +2696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3844,83 +2920,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4145,83 +3144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4446,83 +3368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4747,83 +3592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5048,83 +3816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5349,83 +4040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5650,83 +4264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5951,83 +4488,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6252,83 +4712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6553,83 +4936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6854,83 +5160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7155,83 +5384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7456,83 +5608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7757,83 +5832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8058,83 +6056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8359,83 +6280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8661,83 +6505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8963,83 +6730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9264,83 +6954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9565,83 +7178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9866,83 +7402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10167,83 +7626,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10468,83 +7850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10769,83 +8074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11070,83 +8298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11371,83 +8522,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11672,83 +8746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11973,83 +8970,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12274,83 +9194,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12575,83 +9418,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12876,83 +9642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13177,83 +9866,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13479,83 +10091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13780,83 +10315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14081,83 +10539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14382,83 +10763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14684,83 +10988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14986,83 +11213,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15287,83 +11437,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15588,83 +11661,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15889,83 +11885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16190,83 +12109,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16491,83 +12333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16793,83 +12558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17094,83 +12782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17395,83 +13006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17696,83 +13230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17997,83 +13454,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18298,83 +13678,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18600,83 +13903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18901,83 +14127,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19203,83 +14352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19504,83 +14576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19806,83 +14801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20107,83 +15025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20408,83 +15249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20709,83 +15473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21010,83 +15697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21311,83 +15921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21612,83 +16145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21913,83 +16369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22215,83 +16594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22516,83 +16818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22818,83 +17043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23119,83 +17267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23420,83 +17491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23721,83 +17715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24022,83 +17939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24323,83 +18163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24624,83 +18387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24925,83 +18611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25226,83 +18835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25527,83 +19059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25828,83 +19283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26129,83 +19507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26430,83 +19731,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26732,83 +19956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27034,83 +20181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27336,83 +20406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27637,83 +20630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27938,83 +20854,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28239,83 +21078,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28540,83 +21302,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28841,83 +21526,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29143,83 +21751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29444,83 +21975,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29746,83 +22200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30047,83 +22424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30349,83 +22649,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30651,83 +22874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30952,83 +23098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31253,83 +23322,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31554,83 +23546,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31855,83 +23770,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32156,83 +23994,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32457,83 +24218,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32758,83 +24442,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33059,83 +24666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33361,83 +24891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33663,83 +25116,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33964,83 +25340,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34265,83 +25564,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34566,83 +25788,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34868,83 +26013,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35170,83 +26238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35471,83 +26462,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35772,83 +26686,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36073,83 +26910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36375,83 +27135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36676,83 +27359,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36977,83 +27583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37278,83 +27807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37580,83 +28032,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37882,83 +28257,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38184,83 +28482,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38485,83 +28706,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38786,83 +28930,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39087,83 +29154,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39388,83 +29378,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39689,83 +29602,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39990,83 +29826,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40291,83 +30050,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40592,83 +30274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40893,83 +30498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41194,83 +30722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41495,83 +30946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41796,83 +31170,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42097,83 +31394,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42398,83 +31618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42699,83 +31842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43000,83 +32066,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43302,83 +32291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43603,83 +32515,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43904,83 +32739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44205,83 +32963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44506,83 +33187,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44808,83 +33412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45110,83 +33637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45411,83 +33861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45712,83 +34085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46014,83 +34310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46315,83 +34534,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46617,83 +34759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46919,83 +34984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47220,83 +35208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47521,83 +35432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47822,83 +35656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48123,83 +35880,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48425,83 +36105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48727,83 +36330,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49028,83 +36554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49329,83 +36778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49630,83 +37002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49931,83 +37226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50232,83 +37450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50533,83 +37674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50835,83 +37899,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51136,83 +38123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51437,83 +38347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51738,83 +38571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52039,83 +38795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52341,83 +39020,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52642,83 +39244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52943,83 +39468,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53244,83 +39692,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53546,83 +39917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53847,83 +40141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54148,83 +40365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54449,83 +40589,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54750,83 +40813,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55051,83 +41037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55352,83 +41261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55654,83 +41486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55955,83 +41710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56256,83 +41934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56557,83 +42158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56858,83 +42382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57160,83 +42607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57462,83 +42832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57763,83 +43056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58064,83 +43280,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58365,83 +43504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58666,83 +43728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58967,83 +43952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59268,83 +44176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59569,83 +44400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59870,83 +44624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60171,83 +44848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60472,83 +45072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60773,83 +45296,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61074,83 +45520,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61375,83 +45744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61676,83 +45968,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61977,83 +46192,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62278,83 +46416,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62579,83 +46640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62880,83 +46864,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63181,83 +47088,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63483,83 +47313,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63784,83 +47537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64085,83 +47761,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64386,83 +47985,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64687,83 +48209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64988,83 +48433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65289,83 +48657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65591,83 +48882,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65892,83 +49106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66193,83 +49330,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66494,83 +49554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66796,83 +49779,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67097,83 +50003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67398,83 +50227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67699,83 +50451,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68000,83 +50675,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68301,83 +50899,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68602,83 +51123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68904,83 +51348,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69206,83 +51573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69507,83 +51797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69809,83 +52022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70110,83 +52246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70411,83 +52470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70712,83 +52694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71014,83 +52919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71315,83 +53143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71617,83 +53368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71919,83 +53593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72220,83 +53817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72521,83 +54041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72822,83 +54265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73123,83 +54489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73424,83 +54713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73725,83 +54937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74026,83 +55161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74327,83 +55385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74628,83 +55609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74929,83 +55833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75230,83 +56057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75532,83 +56282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75834,83 +56507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76136,83 +56732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76437,83 +56956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76738,83 +57180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77039,83 +57404,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77340,83 +57628,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77641,83 +57852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77942,83 +58076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78243,83 +58300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78544,83 +58524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78845,83 +58748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79146,83 +58972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79447,83 +59196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79749,83 +59421,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80050,83 +59645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80351,83 +59869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80653,83 +60094,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80954,83 +60318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81255,83 +60542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81557,83 +60767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81859,83 +60992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82160,83 +61216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82461,83 +61440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82762,83 +61664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83063,83 +61888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83364,83 +62112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83666,83 +62337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83967,83 +62561,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84268,83 +62785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84569,83 +63009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84870,83 +63233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85171,83 +63457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85473,83 +63682,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85774,83 +63906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86075,83 +64130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86377,83 +64355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86679,83 +64580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86980,83 +64804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87281,83 +65028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87583,83 +65253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87885,83 +65478,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88186,83 +65702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88488,83 +65927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88789,83 +66151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89090,83 +66375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89391,83 +66599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89693,83 +66824,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89995,83 +67049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90296,83 +67273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90597,83 +67497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90898,83 +67721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91199,83 +67945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91500,83 +68169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91801,83 +68393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92102,83 +68617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92403,83 +68841,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92704,83 +69065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93005,83 +69289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93306,83 +69513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93607,83 +69737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93908,83 +69961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94209,83 +70185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94510,83 +70409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94811,83 +70633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95112,83 +70857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95413,83 +71081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95715,83 +71306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96016,83 +71530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96317,83 +71754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96618,83 +71978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96919,83 +72202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97220,83 +72426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97521,83 +72650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97822,83 +72874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98123,83 +73098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98425,83 +73323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98726,83 +73547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99027,83 +73771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99328,83 +73995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99629,83 +74219,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99930,83 +74443,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100231,83 +74667,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100532,83 +74891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100833,83 +75115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101134,83 +75339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101435,83 +75563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101736,83 +75787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102037,83 +76011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102338,83 +76235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102639,83 +76459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102940,83 +76683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103241,83 +76907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103542,83 +77131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103843,83 +77355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104145,83 +77580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104446,83 +77804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104748,83 +78029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105049,83 +78253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105350,83 +78477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105651,83 +78701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105952,83 +78925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106254,83 +79150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106555,83 +79374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106857,83 +79599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107158,83 +79823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107460,83 +80048,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107762,83 +80273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108064,83 +80498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108366,83 +80723,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108667,83 +80947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108968,83 +81171,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109269,83 +81395,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109570,83 +81619,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109871,83 +81843,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110172,83 +82067,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110473,83 +82291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110774,83 +82515,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111075,83 +82739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111376,83 +82963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111678,83 +83188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111979,83 +83412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112280,83 +83636,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112582,83 +83861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112883,83 +84085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113184,83 +84309,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113485,83 +84533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113786,83 +84757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114087,83 +84981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114388,83 +85205,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114689,83 +85429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114990,83 +85653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115291,83 +85877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115592,83 +86101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115893,83 +86325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116194,83 +86549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116496,83 +86774,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116797,83 +86998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117098,83 +87222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117399,83 +87446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117701,83 +87671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118003,83 +87896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118304,83 +88120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118605,83 +88344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118907,83 +88569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119209,83 +88794,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119510,83 +89018,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119811,83 +89242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120112,83 +89466,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120413,83 +89690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120715,83 +89915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121017,83 +90140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121318,83 +90364,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121619,83 +90588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121921,83 +90813,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122223,83 +91038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122524,83 +91262,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122825,83 +91486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123126,83 +91710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123427,83 +91934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123728,83 +92158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124029,83 +92382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124331,83 +92607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124633,83 +92832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124935,83 +93057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125236,83 +93281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125537,83 +93505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125838,83 +93729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126139,83 +93953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126440,83 +94177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126741,83 +94401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127042,83 +94625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127343,83 +94849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127644,83 +95073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127945,83 +95297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128246,83 +95521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128547,83 +95745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128848,83 +95969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129149,83 +96193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129450,83 +96417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129751,83 +96641,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130052,83 +96865,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130354,83 +97090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130655,83 +97314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130956,83 +97538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131257,83 +97762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131558,83 +97986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131859,83 +98210,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132160,83 +98434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132461,83 +98658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132762,83 +98882,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133064,83 +99107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133366,83 +99332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133667,83 +99556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133968,83 +99780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134269,83 +100004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134570,83 +100228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134871,83 +100452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135172,83 +100676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135473,83 +100900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135774,83 +101124,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136075,83 +101348,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136377,83 +101573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136679,83 +101798,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136980,83 +102022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137281,83 +102246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137582,83 +102470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137883,83 +102694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138184,83 +102918,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138485,83 +103142,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138787,83 +103367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139088,83 +103591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139389,83 +103815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139690,83 +104039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139991,83 +104263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140292,83 +104487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140593,83 +104711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140894,83 +104935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141195,83 +105159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141496,83 +105383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141797,83 +105607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142098,83 +105831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142399,83 +106055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142700,83 +106279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143001,83 +106503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143303,83 +106728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143604,83 +106952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143906,83 +107177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144207,83 +107401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144509,83 +107626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144811,83 +107851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145112,83 +108075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145413,83 +108299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145715,83 +108524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146016,83 +108748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146317,83 +108972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146618,83 +109196,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146919,83 +109420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147220,83 +109644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147521,83 +109868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147822,83 +110092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148123,83 +110316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148424,83 +110540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148726,83 +110765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149028,83 +110990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149329,83 +111214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149631,83 +111439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149932,83 +111663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150233,83 +111887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150534,83 +112111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150836,83 +112336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151137,83 +112560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151438,83 +112784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151740,83 +113009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152041,83 +113233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152342,83 +113457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152643,83 +113681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152944,83 +113905,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153245,83 +114129,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153546,83 +114353,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153847,83 +114577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154148,83 +114801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154450,83 +115026,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154752,83 +115251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155053,83 +115475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155355,83 +115700,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155656,83 +115924,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155957,83 +116148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156258,83 +116372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156560,83 +116597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156861,83 +116821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157162,83 +117045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157463,83 +117269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157765,83 +117494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158066,83 +117718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158368,83 +117943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158669,83 +118167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158970,83 +118391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159272,83 +118616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159573,83 +118840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159874,83 +119064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160175,83 +119288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160476,83 +119512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160777,83 +119736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161078,83 +119960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161379,83 +120184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161680,83 +120408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161981,83 +120632,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162283,83 +120857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162584,83 +121081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162885,83 +121305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163187,83 +121530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163488,83 +121754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163789,83 +121978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164090,83 +122202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164391,83 +122426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164692,83 +122650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164994,83 +122875,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165295,83 +123099,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165596,83 +123323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165897,83 +123547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166198,83 +123771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166499,83 +123995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166800,83 +124219,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167102,83 +124444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167404,83 +124669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167705,83 +124893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168006,83 +125117,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168307,83 +125341,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168608,83 +125565,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168909,83 +125789,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169210,83 +126013,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169512,83 +126238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169813,83 +126462,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170114,83 +126686,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170415,83 +126910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170716,83 +127134,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171017,83 +127358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171318,83 +127582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171619,83 +127806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171920,83 +128030,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172221,83 +128254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172523,83 +128479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172824,83 +128703,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173125,83 +128927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173427,83 +129152,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173729,83 +129377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174030,83 +129601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174331,83 +129825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174632,83 +130049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174934,83 +130274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 7] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 7 - DestDataType: 7 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8HS_BH_BiasSH_HAS_SAB_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8HS_BH_BiasSH_HAS_SAB_SAV_UserArgs.yaml index ff14b170214..605af6505d1 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8HS_BH_BiasSH_HAS_SAB_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_F8HS_BH_BiasSH_HAS_SAB_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -834,83 +680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1135,83 +904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1436,83 +1128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1737,83 +1352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2038,83 +1576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2339,83 +1800,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2640,83 +2024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2941,83 +2248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3242,83 +2472,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3543,83 +2696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3844,83 +2920,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4145,83 +3144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4446,83 +3368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4747,83 +3592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5048,83 +3816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5349,83 +4040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5650,83 +4264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5951,83 +4488,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6252,83 +4712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6553,83 +4936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6854,83 +5160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7155,83 +5384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7456,83 +5608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7757,83 +5832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8058,83 +6056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8359,83 +6280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8661,83 +6505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8963,83 +6730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9264,83 +6954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9565,83 +7178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9866,83 +7402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10167,83 +7626,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10468,83 +7850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10769,83 +8074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11070,83 +8298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11371,83 +8522,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11672,83 +8746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11973,83 +8970,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12274,83 +9194,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12575,83 +9418,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12876,83 +9642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13177,83 +9866,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13479,83 +10091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13780,83 +10315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14081,83 +10539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14382,83 +10763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14684,83 +10988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14986,83 +11213,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15287,83 +11437,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15588,83 +11661,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15889,83 +11885,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16190,83 +12109,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16491,83 +12333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16793,83 +12558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17094,83 +12782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17395,83 +13006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17696,83 +13230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17997,83 +13454,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18298,83 +13678,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18600,83 +13903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18901,83 +14127,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19203,83 +14352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19504,83 +14576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19806,83 +14801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20107,83 +15025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20408,83 +15249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20709,83 +15473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21010,83 +15697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21311,83 +15921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21612,83 +16145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21913,83 +16369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22215,83 +16594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22516,83 +16818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22818,83 +17043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23119,83 +17267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23420,83 +17491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23721,83 +17715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24022,83 +17939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24323,83 +18163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24624,83 +18387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24925,83 +18611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25226,83 +18835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25527,83 +19059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25828,83 +19283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26129,83 +19507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26430,83 +19731,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26732,83 +19956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27034,83 +20181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27336,83 +20406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27637,83 +20630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27938,83 +20854,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28239,83 +21078,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28540,83 +21302,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28841,83 +21526,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29143,83 +21751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29444,83 +21975,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29746,83 +22200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30047,83 +22424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30349,83 +22649,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30651,83 +22874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30952,83 +23098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31253,83 +23322,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31554,83 +23546,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31855,83 +23770,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32156,83 +23994,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32457,83 +24218,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32758,83 +24442,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33059,83 +24666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33361,83 +24891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33663,83 +25116,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33964,83 +25340,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34265,83 +25564,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34566,83 +25788,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34868,83 +26013,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35170,83 +26238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35471,83 +26462,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35772,83 +26686,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36073,83 +26910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36375,83 +27135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36676,83 +27359,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36977,83 +27583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37278,83 +27807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37580,83 +28032,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37882,83 +28257,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38184,83 +28482,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38485,83 +28706,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38786,83 +28930,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39087,83 +29154,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39388,83 +29378,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39689,83 +29602,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39990,83 +29826,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40291,83 +30050,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40592,83 +30274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40893,83 +30498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41194,83 +30722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41495,83 +30946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41796,83 +31170,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42097,83 +31394,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42398,83 +31618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42699,83 +31842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43000,83 +32066,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43302,83 +32291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43603,83 +32515,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43904,83 +32739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44205,83 +32963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44506,83 +33187,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44808,83 +33412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45110,83 +33637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45411,83 +33861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45712,83 +34085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46014,83 +34310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46315,83 +34534,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46617,83 +34759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46919,83 +34984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47220,83 +35208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47521,83 +35432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47822,83 +35656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48123,83 +35880,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48425,83 +36105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48727,83 +36330,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49028,83 +36554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49329,83 +36778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49630,83 +37002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49931,83 +37226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50232,83 +37450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50533,83 +37674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50835,83 +37899,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51136,83 +38123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51437,83 +38347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51738,83 +38571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52039,83 +38795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52341,83 +39020,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52642,83 +39244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52943,83 +39468,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53244,83 +39692,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53546,83 +39917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53847,83 +40141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54148,83 +40365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54449,83 +40589,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54750,83 +40813,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55051,83 +41037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55352,83 +41261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55654,83 +41486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55955,83 +41710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56256,83 +41934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56557,83 +42158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56858,83 +42382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57160,83 +42607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57462,83 +42832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57763,83 +43056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58064,83 +43280,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58365,83 +43504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58666,83 +43728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58967,83 +43952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59268,83 +44176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59569,83 +44400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59870,83 +44624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60171,83 +44848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60472,83 +45072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60773,83 +45296,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61074,83 +45520,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61375,83 +45744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61676,83 +45968,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61977,83 +46192,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62278,83 +46416,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62579,83 +46640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62880,83 +46864,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63181,83 +47088,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63483,83 +47313,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63784,83 +47537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64085,83 +47761,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64386,83 +47985,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64687,83 +48209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64988,83 +48433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65289,83 +48657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65591,83 +48882,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65892,83 +49106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66193,83 +49330,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66494,83 +49554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66796,83 +49779,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67097,83 +50003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67398,83 +50227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67699,83 +50451,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68000,83 +50675,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68301,83 +50899,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68602,83 +51123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68904,83 +51348,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69206,83 +51573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69507,83 +51797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69809,83 +52022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70110,83 +52246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70411,83 +52470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70712,83 +52694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71014,83 +52919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71315,83 +53143,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71617,83 +53368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71919,83 +53593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72220,83 +53817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72521,83 +54041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72822,83 +54265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73123,83 +54489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73424,83 +54713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73725,83 +54937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74026,83 +55161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74327,83 +55385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74628,83 +55609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74929,83 +55833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75230,83 +56057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75532,83 +56282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75834,83 +56507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76136,83 +56732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76437,83 +56956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76738,83 +57180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77039,83 +57404,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77340,83 +57628,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77641,83 +57852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77942,83 +58076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78243,83 +58300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78544,83 +58524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78845,83 +58748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79146,83 +58972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79447,83 +59196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79749,83 +59421,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80050,83 +59645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80351,83 +59869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80653,83 +60094,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80954,83 +60318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81255,83 +60542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81557,83 +60767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81859,83 +60992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82160,83 +61216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82461,83 +61440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82762,83 +61664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83063,83 +61888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83364,83 +62112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83666,83 +62337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83967,83 +62561,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84268,83 +62785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84569,83 +63009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84870,83 +63233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85171,83 +63457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85473,83 +63682,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85774,83 +63906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86075,83 +64130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86377,83 +64355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86679,83 +64580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86980,83 +64804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87281,83 +65028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87583,83 +65253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87885,83 +65478,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88186,83 +65702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88488,83 +65927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88789,83 +66151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89090,83 +66375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89391,83 +66599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89693,83 +66824,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89995,83 +67049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90296,83 +67273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90597,83 +67497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90898,83 +67721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91199,83 +67945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91500,83 +68169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91801,83 +68393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92102,83 +68617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92403,83 +68841,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92704,83 +69065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93005,83 +69289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93306,83 +69513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93607,83 +69737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93908,83 +69961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94209,83 +70185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94510,83 +70409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94811,83 +70633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95112,83 +70857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95413,83 +71081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95715,83 +71306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96016,83 +71530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96317,83 +71754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96618,83 +71978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96919,83 +72202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97220,83 +72426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97521,83 +72650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97822,83 +72874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98123,83 +73098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98425,83 +73323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98726,83 +73547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99027,83 +73771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99328,83 +73995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99629,83 +74219,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99930,83 +74443,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100231,83 +74667,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100532,83 +74891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100833,83 +75115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101134,83 +75339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101435,83 +75563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101736,83 +75787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102037,83 +76011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102338,83 +76235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102639,83 +76459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102940,83 +76683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103241,83 +76907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103542,83 +77131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103843,83 +77355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104145,83 +77580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104446,83 +77804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104748,83 +78029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105049,83 +78253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105350,83 +78477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105651,83 +78701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105952,83 +78925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106254,83 +79150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106555,83 +79374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106857,83 +79599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107158,83 +79823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107460,83 +80048,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107762,83 +80273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108064,83 +80498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108366,83 +80723,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108667,83 +80947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108968,83 +81171,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109269,83 +81395,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109570,83 +81619,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109871,83 +81843,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110172,83 +82067,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110473,83 +82291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110774,83 +82515,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111075,83 +82739,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111376,83 +82963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111678,83 +83188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111979,83 +83412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112280,83 +83636,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112582,83 +83861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112883,83 +84085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113184,83 +84309,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113485,83 +84533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113786,83 +84757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114087,83 +84981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114388,83 +85205,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114689,83 +85429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114990,83 +85653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115291,83 +85877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115592,83 +86101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115893,83 +86325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116194,83 +86549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116496,83 +86774,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116797,83 +86998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117098,83 +87222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117399,83 +87446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117701,83 +87671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118003,83 +87896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118304,83 +88120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118605,83 +88344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118907,83 +88569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119209,83 +88794,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119510,83 +89018,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119811,83 +89242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120112,83 +89466,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120413,83 +89690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120715,83 +89915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121017,83 +90140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121318,83 +90364,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121619,83 +90588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121921,83 +90813,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122223,83 +91038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122524,83 +91262,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122825,83 +91486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123126,83 +91710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123427,83 +91934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123728,83 +92158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124029,83 +92382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124331,83 +92607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124633,83 +92832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124935,83 +93057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125236,83 +93281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125537,83 +93505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125838,83 +93729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126139,83 +93953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126440,83 +94177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126741,83 +94401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127042,83 +94625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127343,83 +94849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127644,83 +95073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127945,83 +95297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128246,83 +95521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128547,83 +95745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128848,83 +95969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129149,83 +96193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129450,83 +96417,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129751,83 +96641,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130052,83 +96865,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130354,83 +97090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130655,83 +97314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130956,83 +97538,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131257,83 +97762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131558,83 +97986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131859,83 +98210,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132160,83 +98434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132461,83 +98658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132762,83 +98882,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133064,83 +99107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133366,83 +99332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133667,83 +99556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133968,83 +99780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134269,83 +100004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134570,83 +100228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134871,83 +100452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135172,83 +100676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135473,83 +100900,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135774,83 +101124,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136075,83 +101348,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136377,83 +101573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136679,83 +101798,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136980,83 +102022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137281,83 +102246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137582,83 +102470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137883,83 +102694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138184,83 +102918,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138485,83 +103142,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138787,83 +103367,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139088,83 +103591,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139389,83 +103815,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139690,83 +104039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139991,83 +104263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140292,83 +104487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140593,83 +104711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140894,83 +104935,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141195,83 +105159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141496,83 +105383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141797,83 +105607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142098,83 +105831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142399,83 +106055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142700,83 +106279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143001,83 +106503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143303,83 +106728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143604,83 +106952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143906,83 +107177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144207,83 +107401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144509,83 +107626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144811,83 +107851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145112,83 +108075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145413,83 +108299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145715,83 +108524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146016,83 +108748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146317,83 +108972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146618,83 +109196,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146919,83 +109420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147220,83 +109644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147521,83 +109868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147822,83 +110092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148123,83 +110316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148424,83 +110540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148726,83 +110765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149028,83 +110990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149329,83 +111214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149631,83 +111439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149932,83 +111663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150233,83 +111887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150534,83 +112111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150836,83 +112336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151137,83 +112560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151438,83 +112784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151740,83 +113009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152041,83 +113233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152342,83 +113457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152643,83 +113681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152944,83 +113905,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153245,83 +114129,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153546,83 +114353,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153847,83 +114577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154148,83 +114801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154450,83 +115026,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154752,83 +115251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155053,83 +115475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155355,83 +115700,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155656,83 +115924,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155957,83 +116148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156258,83 +116372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156560,83 +116597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156861,83 +116821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157162,83 +117045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157463,83 +117269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157765,83 +117494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158066,83 +117718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158368,83 +117943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158669,83 +118167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158970,83 +118391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159272,83 +118616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159573,83 +118840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159874,83 +119064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160175,83 +119288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160476,83 +119512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160777,83 +119736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161078,83 +119960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161379,83 +120184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161680,83 +120408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161981,83 +120632,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162283,83 +120857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162584,83 +121081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162885,83 +121305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163187,83 +121530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163488,83 +121754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163789,83 +121978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164090,83 +122202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164391,83 +122426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164692,83 +122650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164994,83 +122875,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165295,83 +123099,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165596,83 +123323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165897,83 +123547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166198,83 +123771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166499,83 +123995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166800,83 +124219,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167102,83 +124444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167404,83 +124669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167705,83 +124893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168006,83 +125117,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168307,83 +125341,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168608,83 +125565,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168909,83 +125789,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169210,83 +126013,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169512,83 +126238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169813,83 +126462,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170114,83 +126686,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170415,83 +126910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170716,83 +127134,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171017,83 +127358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171318,83 +127582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171619,83 +127806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171920,83 +128030,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172221,83 +128254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172523,83 +128479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172824,83 +128703,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173125,83 +128927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173427,83 +129152,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173729,83 +129377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174030,83 +129601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174331,83 +129825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174632,83 +130049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174934,83 +130274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 15 - DataTypeA: 15 - DataTypeAmaxD: 0 - DataTypeB: 15 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: Scalar - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_Aux_AH_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_Aux_AH_SAV.yaml index 8a8b584dbc7..e9ffaf22612 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_Aux_AH_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_Aux_AH_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -401,66 +341,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -625,66 +505,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcA_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcA_SAV.yaml index 271ff20b4eb..6146a4e3fcf 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcA_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcA_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: A - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -400,66 +340,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: A - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -623,66 +503,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: A - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcB_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcB_SAV.yaml index 22e8437de2f..ac0af75532e 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcB_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcB_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -400,66 +340,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -623,66 +503,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml index 3c2607b1188..efb9e1033d2 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -401,66 +341,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_GG_AS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_GG_AS_SAV_UserArgs.yaml index 88ee999cde8..9386d361d83 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_GG_AS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_GG_AS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -70,7 +70,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 UseScaleCD: false - - 1LDSBuffer: 0 @@ -202,78 +202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: true - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml index 342be4f65d3..b88f152e80c 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HHS_BH_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -222,81 +222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -506,81 +431,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -790,81 +640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1074,81 +849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1358,81 +1058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1642,81 +1267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1926,81 +1476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2210,81 +1685,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2494,81 +1894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2778,81 +2103,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3062,81 +2312,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3346,81 +2521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3630,81 +2730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3914,81 +2939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4198,81 +3148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4482,81 +3357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4766,81 +3566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5050,81 +3775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5334,81 +3984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5618,81 +4193,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5902,81 +4402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6186,81 +4611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6470,81 +4820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6754,81 +5029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7038,81 +5238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7322,81 +5447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7606,81 +5656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7890,81 +5865,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8174,81 +6074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8458,81 +6283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8742,81 +6492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9026,81 +6701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9310,81 +6910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9594,81 +7119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9878,81 +7328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10162,81 +7537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10446,81 +7746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10730,81 +7955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11014,81 +8164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11298,81 +8373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11582,81 +8582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11866,81 +8791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12150,81 +9000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12434,81 +9209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12718,81 +9418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13002,81 +9627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13286,81 +9836,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13570,81 +10045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13854,81 +10254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14138,81 +10463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14422,81 +10672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14706,81 +10881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14990,81 +11090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15274,81 +11299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15558,81 +11508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15842,81 +11717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16126,81 +11926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16410,81 +12135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16694,81 +12344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16978,81 +12553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17262,81 +12762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17546,81 +12971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17830,81 +13180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18114,81 +13389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18398,81 +13598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18682,81 +13807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18966,81 +14016,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19250,81 +14225,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19534,81 +14434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19818,81 +14643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20102,81 +14852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20386,81 +15061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20670,81 +15270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20954,81 +15479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21238,81 +15688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21522,81 +15897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21806,81 +16106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22090,81 +16315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22374,81 +16524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22658,81 +16733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22942,81 +16942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23226,81 +17151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23510,81 +17360,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23794,81 +17569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24078,81 +17778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24362,81 +17987,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24646,81 +18196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24930,81 +18405,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25214,81 +18614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25498,81 +18823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25782,81 +19032,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26074,83 +19249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26375,83 +19473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26676,83 +19697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26977,83 +19921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27278,83 +20145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27579,83 +20369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27880,83 +20593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28181,83 +20817,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28482,83 +21041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28783,83 +21265,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29084,83 +21489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29385,83 +21713,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29686,83 +21937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29987,83 +22161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30288,83 +22385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30590,83 +22610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30892,83 +22835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31193,83 +23059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31494,83 +23283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31795,83 +23507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32097,83 +23732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32399,83 +23957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32701,83 +24182,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33002,83 +24406,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33304,83 +24631,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33606,83 +24856,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33907,83 +25080,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34208,83 +25304,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34509,83 +25528,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34810,83 +25752,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35111,83 +25976,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35413,83 +26201,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35714,83 +26425,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36015,83 +26649,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36317,83 +26874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36618,83 +27098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36919,83 +27322,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37220,83 +27546,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37522,83 +27771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37824,83 +27996,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38125,83 +28220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38426,83 +28444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38727,83 +28668,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39028,83 +28892,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39329,83 +29116,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39630,83 +29340,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39931,83 +29564,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40232,83 +29788,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40533,83 +30012,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40834,83 +30236,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41135,83 +30460,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41437,83 +30685,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41739,83 +30910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42041,83 +31135,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42342,83 +31359,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42643,83 +31583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42944,83 +31807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43246,83 +32032,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43547,83 +32256,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43848,83 +32480,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44149,83 +32704,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44450,83 +32928,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44752,83 +33153,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45053,83 +33377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45354,83 +33601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45656,83 +33826,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45958,83 +34051,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46259,83 +34275,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46560,83 +34499,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46861,83 +34723,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47162,83 +34947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47464,83 +35172,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47765,83 +35396,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48066,83 +35620,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48367,83 +35844,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48669,83 +36069,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48970,83 +36293,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49271,83 +36517,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49572,83 +36741,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49873,83 +36965,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50174,83 +37189,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50475,83 +37413,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50776,83 +37637,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51078,83 +37862,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51379,83 +38086,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51680,83 +38310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51982,83 +38535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52284,83 +38760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52585,83 +38984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52887,83 +39209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53188,83 +39433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53489,83 +39657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53790,83 +39881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54092,83 +40106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54393,83 +40330,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54694,83 +40554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54995,83 +40778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55296,83 +41002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55597,83 +41226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55898,83 +41450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56199,83 +41674,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56501,83 +41899,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56802,83 +42123,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57103,83 +42347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57404,83 +42571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57706,83 +42796,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58008,83 +43021,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58310,83 +43246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58611,83 +43470,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58912,83 +43694,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59214,83 +43919,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59516,83 +44144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59817,83 +44368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60118,83 +44592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60420,83 +44817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60721,83 +45041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61023,83 +45266,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61324,83 +45490,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61625,83 +45714,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61926,83 +45938,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62227,83 +46162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62528,83 +46386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62829,83 +46610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63130,83 +46834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63431,83 +47058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63732,83 +47282,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64034,83 +47507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64335,83 +47731,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64636,83 +47955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64937,83 +48179,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65238,83 +48403,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65539,83 +48627,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65840,83 +48851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66141,83 +49075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66443,83 +49300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66744,83 +49524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67045,83 +49748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67346,83 +49972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67647,83 +50196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67948,83 +50420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68250,83 +50645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68552,83 +50870,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68853,83 +51094,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69154,83 +51318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69455,83 +51542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69756,83 +51766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70057,83 +51990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70358,83 +52214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70659,83 +52438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70960,83 +52662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71261,83 +52886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71562,83 +53110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71863,83 +53334,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72164,83 +53558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72465,83 +53782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72766,83 +54006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73067,83 +54230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73368,83 +54454,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73669,83 +54678,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73971,83 +54903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74273,83 +55128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74574,83 +55352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74875,83 +55576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75176,83 +55800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75477,83 +56024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75778,83 +56248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76079,83 +56472,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76380,83 +56696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76681,83 +56920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76982,83 +57144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77283,83 +57368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77584,83 +57592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77885,83 +57816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78186,83 +58040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78487,83 +58264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78788,83 +58488,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79089,83 +58712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79391,83 +58937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79692,83 +59161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79993,83 +59385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80294,83 +59609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80595,83 +59833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80896,83 +60057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81197,83 +60281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81498,83 +60505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81799,83 +60729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82100,83 +60953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82401,83 +61177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82702,83 +61401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83003,83 +61625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83304,83 +61849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83605,83 +62073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83906,83 +62297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84207,83 +62521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84508,83 +62745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84810,83 +62970,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85112,83 +63195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85414,83 +63420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85715,83 +63644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86016,83 +63868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86318,83 +64093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86620,83 +64318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86921,83 +64542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87223,83 +64767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87525,83 +64992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87826,83 +65216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88127,83 +65440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88428,83 +65664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88729,83 +65888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89030,83 +66112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89332,83 +66337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89633,83 +66561,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89934,83 +66785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90235,83 +67009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90536,83 +67233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90838,83 +67458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91139,83 +67682,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91440,83 +67906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91741,83 +68130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92043,83 +68355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92344,83 +68579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92645,83 +68803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92947,83 +69028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93248,83 +69252,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93550,83 +69477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93851,83 +69701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94152,83 +69925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94453,83 +70149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94755,83 +70374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95056,83 +70598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95358,83 +70823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95659,83 +71047,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95960,83 +71271,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96262,83 +71496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96564,83 +71721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96865,83 +71945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97166,83 +72169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97467,83 +72393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97768,83 +72617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98070,83 +72842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98372,83 +73067,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98673,83 +73291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98975,83 +73516,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99276,83 +73740,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99578,83 +73965,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99880,83 +74190,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100181,83 +74414,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100482,83 +74638,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100783,83 +74862,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101085,83 +75087,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101386,83 +75311,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101687,83 +75535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101988,83 +75759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102289,83 +75983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102590,83 +76207,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102892,83 +76432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103193,83 +76656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103495,83 +76881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103796,83 +77105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104097,83 +77329,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104398,83 +77553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104700,83 +77778,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105002,83 +78003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105304,83 +78228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105605,83 +78452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105906,83 +78676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106208,83 +78901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106509,83 +79125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106811,83 +79350,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107112,83 +79574,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107413,83 +79798,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107714,83 +80022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108015,83 +80246,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108317,83 +80471,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108619,83 +80696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108921,83 +80921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109223,83 +81146,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109525,83 +81371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109827,83 +81596,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110129,83 +81821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110430,83 +82045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110731,83 +82269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111032,83 +82493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111333,83 +82717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111634,83 +82941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111936,83 +83166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112237,83 +83390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112538,83 +83614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112839,83 +83838,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113141,83 +84063,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113443,83 +84288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113744,83 +84512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114046,83 +84737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114347,83 +84961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114648,83 +85185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114950,83 +85410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115252,83 +85635,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115554,83 +85860,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115855,83 +86084,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116156,83 +86308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116457,83 +86532,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116758,83 +86756,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117059,83 +86980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117361,83 +87205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117662,83 +87429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117963,83 +87653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118264,83 +87877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118565,83 +88101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118866,83 +88325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119168,83 +88550,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119470,83 +88775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119771,83 +88999,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120073,83 +89224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120375,83 +89449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120677,83 +89674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120978,83 +89898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121279,83 +90122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121581,83 +90347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121882,83 +90571,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122184,83 +90796,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122485,83 +91020,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122786,83 +91244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123087,83 +91468,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123389,83 +91693,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123690,83 +91917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123991,83 +92141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124293,83 +92366,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124594,83 +92590,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124895,83 +92814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125196,83 +93038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125497,83 +93262,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125798,83 +93486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126099,83 +93710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126400,83 +93934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126701,83 +94158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127002,83 +94382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127303,83 +94606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127605,83 +94831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127907,83 +95056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128208,83 +95280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128509,83 +95504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128810,83 +95728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129111,83 +95952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129412,83 +96176,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129713,83 +96400,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130014,83 +96624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130315,83 +96848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130616,83 +97072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130917,83 +97296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131218,83 +97520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131520,83 +97745,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131821,83 +97969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132123,83 +98194,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132424,83 +98418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132725,83 +98642,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133026,83 +98866,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133327,83 +99090,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133629,83 +99315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133930,83 +99539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134232,83 +99764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134533,83 +99988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134834,83 +100212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135135,83 +100436,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135436,83 +100660,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135737,83 +100884,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136038,83 +101108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136339,83 +101332,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136640,83 +101556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136941,83 +101780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137242,83 +102004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137544,83 +102229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137845,83 +102453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138146,83 +102677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138447,83 +102901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138748,83 +103125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139050,83 +103350,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139352,83 +103575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139653,83 +103799,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139955,83 +104024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140256,83 +104248,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140558,83 +104473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140859,83 +104697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141160,83 +104921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141461,83 +105145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141762,83 +105369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142063,83 +105593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142365,83 +105818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142666,83 +106042,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142968,83 +106267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143269,83 +106491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143570,83 +106715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143871,83 +106939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144173,83 +107164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144474,83 +107388,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144775,83 +107612,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145077,83 +107837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145378,83 +108061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145680,83 +108286,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145981,83 +108510,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146282,83 +108734,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146584,83 +108959,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146885,83 +109183,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147186,83 +109407,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147487,83 +109631,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147789,83 +109856,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148090,83 +110080,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148391,83 +110304,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148692,83 +110528,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148993,83 +110752,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149294,83 +110976,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149595,83 +111200,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149897,83 +111425,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150198,83 +111649,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150499,83 +111873,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150801,83 +112098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151103,83 +112323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151404,83 +112547,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151705,83 +112771,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152006,83 +112995,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152308,83 +113220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152609,83 +113444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152910,83 +113668,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153212,83 +113893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153514,83 +114118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153815,83 +114342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154116,83 +114566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154417,83 +114790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154718,83 +115014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155019,83 +115238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155320,83 +115462,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155621,83 +115686,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155922,83 +115910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156223,83 +116134,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156524,83 +116358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156825,83 +116582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157126,83 +116806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157427,83 +117030,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157728,83 +117254,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158029,83 +117478,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158331,83 +117703,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158632,83 +117927,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158933,83 +118151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159234,83 +118375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159536,83 +118600,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159838,83 +118825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160139,83 +119049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160441,83 +119274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160742,83 +119498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161043,83 +119722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161345,83 +119947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161646,83 +120171,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161947,83 +120395,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162249,83 +120620,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162550,83 +120844,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162852,83 +121069,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163154,83 +121294,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163455,83 +121518,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163756,83 +121742,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164057,83 +121966,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164359,83 +122191,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164660,83 +122415,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164961,83 +122639,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165262,83 +122863,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165563,83 +123087,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165864,83 +123311,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166165,83 +123535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166466,83 +123759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166767,83 +123983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167068,83 +124207,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167369,83 +124431,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167670,83 +124655,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167971,83 +124879,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168272,83 +125103,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168574,83 +125328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168876,83 +125553,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169177,83 +125777,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169479,83 +126002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169780,83 +126226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170081,83 +126450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170382,83 +126674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170683,83 +126898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170984,83 +127122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171285,83 +127346,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171586,83 +127570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171887,83 +127794,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172188,83 +128018,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172489,83 +128242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172790,83 +128466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173091,83 +128690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173392,83 +128914,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173693,83 +129138,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173994,83 +129362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174296,83 +129587,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174597,83 +129811,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174898,83 +130035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175200,83 +130260,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175502,83 +130485,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175804,83 +130710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176105,83 +130934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176406,83 +131158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176708,83 +131383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177010,83 +131608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177312,83 +131833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177613,83 +132057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177914,83 +132281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178216,83 +132506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178517,83 +132730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178818,83 +132954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179120,83 +133179,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179421,83 +133403,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179722,83 +133627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180024,83 +133852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180325,83 +134076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180626,83 +134300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180928,83 +134525,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181229,83 +134749,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181530,83 +134973,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181831,83 +135197,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182132,83 +135421,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182433,83 +135645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182734,83 +135869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183035,83 +136093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183336,83 +136317,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183637,83 +136541,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183939,83 +136766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184241,83 +136991,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184542,83 +137215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184843,83 +137439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185144,83 +137663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185445,83 +137887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185746,83 +138111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186048,83 +138336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186349,83 +138560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186650,83 +138784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186951,83 +139008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187253,83 +139233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187554,83 +139457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187855,83 +139681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188157,83 +139906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188458,83 +140130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188759,83 +140354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189060,83 +140578,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189362,83 +140803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189663,83 +141027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189965,83 +141252,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190266,83 +141476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190568,83 +141701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190869,83 +141925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191170,83 +142149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191471,83 +142373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191772,83 +142597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192073,83 +142821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192375,83 +143046,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192676,83 +143270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192978,83 +143495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193279,83 +143719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193580,83 +143943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193881,83 +144167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194182,83 +144391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194484,83 +144616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194785,83 +144840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195086,83 +145064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195387,83 +145288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195689,83 +145513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195990,83 +145737,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196291,83 +145961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196592,83 +146185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196893,83 +146409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197194,83 +146633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197495,83 +146857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197796,83 +147081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198097,83 +147305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198399,83 +147530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198700,83 +147754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199001,83 +147978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199302,83 +148202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199603,83 +148426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199904,83 +148650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200205,83 +148874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200506,83 +149098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200807,83 +149322,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201108,83 +149546,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201409,83 +149770,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201710,83 +149994,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202011,83 +150218,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202312,83 +150442,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202613,83 +150666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202914,83 +150890,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203215,83 +151114,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203516,83 +151338,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203817,83 +151562,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204119,83 +151787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204421,83 +152012,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204722,83 +152236,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205023,83 +152460,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205324,83 +152684,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205625,83 +152908,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205927,83 +153133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206228,83 +153357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206529,83 +153581,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206831,83 +153806,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207133,83 +154031,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207434,83 +154255,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207736,83 +154480,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208037,83 +154704,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208339,83 +154929,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208640,83 +155153,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208941,83 +155377,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209242,83 +155601,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209543,83 +155825,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209844,83 +156049,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210146,83 +156274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210447,83 +156498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210748,83 +156722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211049,83 +156946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211350,83 +157170,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211651,83 +157394,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211952,83 +157618,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212253,83 +157842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212554,83 +158066,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212855,83 +158290,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213156,83 +158514,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213457,83 +158738,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213758,83 +158962,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214060,83 +159187,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214361,83 +159411,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214663,83 +159636,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214965,83 +159861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215266,83 +160085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215567,83 +160309,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215868,83 +160533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216169,83 +160757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216471,83 +160982,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216772,83 +161206,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217073,83 +161430,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217374,83 +161654,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217675,83 +161878,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217976,83 +162102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218278,83 +162327,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218579,83 +162551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218880,83 +162775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219181,83 +162999,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219482,83 +163223,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219783,83 +163447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220084,83 +163671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220385,83 +163895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220687,83 +164120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220989,83 +164345,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221291,83 +164570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221593,83 +164795,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221894,83 +165019,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222195,83 +165243,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222496,83 +165467,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222798,83 +165692,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223099,83 +165916,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223400,83 +166140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223701,83 +166364,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224002,83 +166588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224303,83 +166812,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224604,83 +167036,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224905,83 +167260,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225207,83 +167485,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225508,83 +167709,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225809,83 +167933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226111,83 +168158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226412,83 +168382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226713,83 +168606,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227014,83 +168830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227315,83 +169054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227616,83 +169278,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227917,83 +169502,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228218,83 +169726,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228519,83 +169950,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228820,83 +170174,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229121,83 +170398,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229423,83 +170623,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229725,83 +170848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230026,83 +171072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230328,83 +171297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230630,83 +171522,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230932,83 +171747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231233,83 +171971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231534,83 +172195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231835,83 +172419,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232136,83 +172643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232437,83 +172867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232738,83 +173091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233040,83 +173316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233341,83 +173540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233643,83 +173765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233944,83 +173989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234246,83 +174214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234548,83 +174439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234849,83 +174663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235151,83 +174888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235453,83 +175113,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235754,83 +175337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236055,83 +175561,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236357,83 +175786,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236659,83 +176011,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236960,83 +176235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237261,83 +176459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237562,83 +176683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237863,83 +176907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238165,83 +177132,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238467,83 +177357,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238768,83 +177581,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239069,83 +177805,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239370,83 +178029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239671,83 +178253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239972,83 +178477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240273,83 +178701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240574,83 +178925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240875,83 +179149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241176,83 +179373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241477,83 +179597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241779,83 +179822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242080,83 +180046,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242381,83 +180270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242682,83 +180494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242983,83 +180718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243284,83 +180942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243585,83 +181166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243886,83 +181390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244187,83 +181614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244488,83 +181838,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244789,83 +182062,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245091,83 +182287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245393,83 +182512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245694,83 +182736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245996,83 +182961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246298,83 +183186,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246599,83 +183410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246900,83 +183634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247201,83 +183858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247502,83 +184082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247803,83 +184306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248104,83 +184530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248406,83 +184755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248707,83 +184979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249009,83 +185204,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249310,83 +185428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249611,83 +185652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249912,83 +185876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250214,83 +186101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250515,83 +186325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250816,83 +186549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251117,83 +186773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251419,83 +186998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251721,83 +187223,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252022,83 +187447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 4 - DestDataType: 4 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HSS_BH_Bias_GG_AS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HSS_BH_Bias_GG_AS_SAV_UserArgs.yaml index 0c497cb8b66..d640d6815f9 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HSS_BH_Bias_GG_AS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HSS_BH_Bias_GG_AS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -70,7 +70,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 UseScaleCD: false - - 1LDSBuffer: 0 @@ -202,78 +202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: true - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HSS_BH_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HSS_BH_Bias_HAS_SAV_UserArgs.yaml index 6a91de39ecc..b35183ba049 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HSS_BH_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_HSS_BH_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -834,83 +680,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1135,83 +904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1436,83 +1128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1738,83 +1353,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2039,83 +1577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2340,83 +1801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2641,83 +2025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2942,83 +2249,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3243,83 +2473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3544,83 +2697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3845,83 +2921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4146,83 +3145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4447,83 +3369,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4748,83 +3593,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5049,83 +3817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5350,83 +4041,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5651,83 +4265,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5952,83 +4489,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6253,83 +4713,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6554,83 +4937,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6855,83 +5161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7156,83 +5385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7457,83 +5609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7759,83 +5834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8060,83 +6058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8361,83 +6282,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8662,83 +6506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8963,83 +6730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9265,83 +6955,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9567,83 +7180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9868,83 +7404,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10169,83 +7628,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10471,83 +7853,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10772,83 +8077,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11073,83 +8301,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11374,83 +8525,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11676,83 +8750,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11977,83 +8974,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12279,83 +9199,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12580,83 +9423,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12881,83 +9647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13182,83 +9871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13484,83 +10096,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13785,83 +10320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14086,83 +10544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14387,83 +10768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14688,83 +10992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14989,83 +11216,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15290,83 +11440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15591,83 +11664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15893,83 +11889,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16194,83 +12113,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16495,83 +12337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16796,83 +12561,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17097,83 +12785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17398,83 +13009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17699,83 +13233,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18000,83 +13457,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18302,83 +13682,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18603,83 +13906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18905,83 +14131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19206,83 +14355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19507,83 +14579,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19808,83 +14803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20109,83 +15027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20410,83 +15251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20711,83 +15475,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21012,83 +15699,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21313,83 +15923,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21614,83 +16147,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21915,83 +16371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22216,83 +16595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22517,83 +16819,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22818,83 +17043,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23119,83 +17267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23420,83 +17491,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23721,83 +17715,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24023,83 +17940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24325,83 +18165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24627,83 +18390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24929,83 +18615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25231,83 +18840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25532,83 +19064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25833,83 +19288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26134,83 +19512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26436,83 +19737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26738,83 +19962,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27039,83 +20186,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27340,83 +20410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27642,83 +20635,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27943,83 +20859,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28244,83 +21083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28545,83 +21307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28847,83 +21532,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29148,83 +21756,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29450,83 +21981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29751,83 +22205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30052,83 +22429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30354,83 +22654,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30656,83 +22879,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30957,83 +23103,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31258,83 +23327,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31559,83 +23551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31860,83 +23775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32161,83 +23999,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32463,83 +24224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32765,83 +24449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33066,83 +24673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33367,83 +24897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33669,83 +25122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33970,83 +25346,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34271,83 +25570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34572,83 +25794,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34874,83 +26019,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35175,83 +26243,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35476,83 +26467,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35777,83 +26691,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36078,83 +26915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36380,83 +27140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36681,83 +27364,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36982,83 +27588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37284,83 +27813,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37586,83 +28038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37887,83 +28262,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38189,83 +28487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38491,83 +28712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38792,83 +28936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39093,83 +29160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39394,83 +29384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39695,83 +29608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39996,83 +29832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40298,83 +30057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40599,83 +30281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40900,83 +30505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41201,83 +30729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41503,83 +30954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41804,83 +31178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42105,83 +31402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42407,83 +31627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42708,83 +31851,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43009,83 +32075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43310,83 +32299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43611,83 +32523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43912,83 +32747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44213,83 +32971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44514,83 +33195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44815,83 +33419,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45116,83 +33643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45418,83 +33868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45719,83 +34092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46020,83 +34316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46322,83 +34541,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46624,83 +34766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46925,83 +34990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47227,83 +35215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47529,83 +35440,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47830,83 +35664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48131,83 +35888,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48432,83 +36112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48733,83 +36336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49034,83 +36560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49335,83 +36784,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49637,83 +37009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49939,83 +37234,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50240,83 +37458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50542,83 +37683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50843,83 +37907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51145,83 +38132,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51446,83 +38356,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51748,83 +38581,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52049,83 +38805,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52350,83 +39029,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52651,83 +39253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52952,83 +39477,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53253,83 +39701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53554,83 +39925,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53855,83 +40149,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54157,83 +40374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54458,83 +40598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54759,83 +40822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55060,83 +41046,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55362,83 +41271,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55663,83 +41495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55965,83 +41720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56266,83 +41944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56567,83 +42168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56868,83 +42392,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57170,83 +42617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57471,83 +42841,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57772,83 +43065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58073,83 +43289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58374,83 +43513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58676,83 +43738,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58978,83 +43963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59280,83 +44188,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59581,83 +44412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59882,83 +44636,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60184,83 +44861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60485,83 +45085,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60786,83 +45309,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61088,83 +45534,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61389,83 +45758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61691,83 +45983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61992,83 +46207,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62293,83 +46431,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62594,83 +46655,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62895,83 +46879,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63196,83 +47103,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63497,83 +47327,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63798,83 +47551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64100,83 +47776,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64402,83 +48001,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64704,83 +48226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65005,83 +48450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65306,83 +48674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65607,83 +48898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65908,83 +49122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66210,83 +49347,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66512,83 +49572,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66813,83 +49796,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67114,83 +50020,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67415,83 +50244,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67716,83 +50468,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68017,83 +50692,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68319,83 +50917,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68620,83 +51141,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68922,83 +51366,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69223,83 +51590,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69524,83 +51814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69826,83 +52039,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70127,83 +52263,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70428,83 +52487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70729,83 +52711,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71031,83 +52936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71333,83 +53161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71634,83 +53385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71935,83 +53609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72236,83 +53833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72537,83 +54057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72838,83 +54281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73140,83 +54506,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73441,83 +54730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73742,83 +54954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74043,83 +55178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74344,83 +55402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74646,83 +55627,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74948,83 +55852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75249,83 +56076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75550,83 +56300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75852,83 +56525,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76153,83 +56749,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76454,83 +56973,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76755,83 +57197,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77057,83 +57422,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77358,83 +57646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77659,83 +57870,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77961,83 +58095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78262,83 +58319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78564,83 +58544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78865,83 +58768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79166,83 +58992,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79468,83 +59217,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79770,83 +59442,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80071,83 +59666,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80373,83 +59891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80674,83 +60115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80976,83 +60340,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81278,83 +60565,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81579,83 +60789,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81881,83 +61014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82183,83 +61239,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82485,83 +61464,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82786,83 +61688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83087,83 +61912,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83388,83 +62136,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83689,83 +62360,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83991,83 +62585,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84293,83 +62810,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84595,83 +63035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84897,83 +63260,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85198,83 +63484,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85500,83 +63709,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85802,83 +63934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86103,83 +64158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86405,83 +64383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86707,83 +64608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87008,83 +64832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87309,83 +65056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87610,83 +65280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87911,83 +65504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88212,83 +65728,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88513,83 +65952,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88815,83 +66177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89117,83 +66402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89418,83 +66626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89719,83 +66850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90020,83 +67074,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90321,83 +67298,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90623,83 +67523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90925,83 +67748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91226,83 +67972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91527,83 +68196,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91828,83 +68420,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92129,83 +68644,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92430,83 +68868,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92732,83 +69093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93034,83 +69318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93335,83 +69542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93636,83 +69766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93937,83 +69990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94238,83 +70214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94539,83 +70438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94840,83 +70662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95141,83 +70886,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95442,83 +71110,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95743,83 +71334,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96044,83 +71558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96345,83 +71782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96647,83 +72007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96948,83 +72231,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97249,83 +72455,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97551,83 +72680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97853,83 +72905,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98155,83 +73130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98457,83 +73355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98759,83 +73580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99060,83 +73804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99361,83 +74028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99663,83 +74253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99964,83 +74477,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100265,83 +74701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100567,83 +74926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100868,83 +75150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101169,83 +75374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101470,83 +75598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101771,83 +75822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102072,83 +76046,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102373,83 +76270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102674,83 +76494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102975,83 +76718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103276,83 +76942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103577,83 +77166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103879,83 +77391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104181,83 +77616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104482,83 +77840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104783,83 +78064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105084,83 +78288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105385,83 +78512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105687,83 +78737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105989,83 +78962,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106291,83 +79187,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106593,83 +79412,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106894,83 +79636,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107196,83 +79861,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107498,83 +80086,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107799,83 +80310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108100,83 +80534,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108401,83 +80758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108703,83 +80983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109004,83 +81207,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109305,83 +81431,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109606,83 +81655,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109908,83 +81880,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110209,83 +82104,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110510,83 +82328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110811,83 +82552,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111113,83 +82777,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111415,83 +83002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111716,83 +83226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112017,83 +83450,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112318,83 +83674,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112619,83 +83898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112920,83 +84122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113221,83 +84346,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113522,83 +84570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113823,83 +84794,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114124,83 +85018,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114425,83 +85242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114726,83 +85466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115027,83 +85690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115329,83 +85915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115630,83 +86139,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115931,83 +86363,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116232,83 +86587,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116533,83 +86811,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116834,83 +87035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117135,83 +87259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117436,83 +87483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117738,83 +87708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118039,83 +87932,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118340,83 +88156,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118642,83 +88381,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118943,83 +88605,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119244,83 +88829,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119545,83 +89053,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119846,83 +89277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120148,83 +89502,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120449,83 +89726,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120750,83 +89950,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121051,83 +90174,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121352,83 +90398,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121654,83 +90623,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121955,83 +90847,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122257,83 +91072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122558,83 +91296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122859,83 +91520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123160,83 +91744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123461,83 +91968,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123762,83 +92192,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124063,83 +92416,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124364,83 +92640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124665,83 +92864,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124966,83 +93088,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125267,83 +93312,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125568,83 +93536,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125870,83 +93761,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126171,83 +93985,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126472,83 +94209,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126774,83 +94434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127076,83 +94659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127377,83 +94883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127678,83 +95107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127979,83 +95331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128280,83 +95555,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128581,83 +95779,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128882,83 +96003,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129184,83 +96228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129485,83 +96452,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129787,83 +96677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130089,83 +96902,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130390,83 +97126,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130692,83 +97351,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130993,83 +97575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131294,83 +97799,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131595,83 +98023,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131896,83 +98247,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132197,83 +98471,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132498,83 +98695,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132800,83 +98920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133101,83 +99144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133402,83 +99368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133703,83 +99592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134005,83 +99817,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134306,83 +100041,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134608,83 +100266,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134909,83 +100490,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135210,83 +100714,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135511,83 +100938,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135812,83 +101162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136113,83 +101386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136414,83 +101610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136715,83 +101834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137017,83 +102059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137319,83 +102284,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137621,83 +102509,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137922,83 +102733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138223,83 +102957,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138524,83 +103181,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138826,83 +103406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139127,83 +103630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139429,83 +103855,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139731,83 +104080,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140032,83 +104304,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140333,83 +104528,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140635,83 +104753,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140937,83 +104978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141239,83 +105203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141540,83 +105427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141841,83 +105651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142143,83 +105876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142444,83 +106100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142746,83 +106325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143047,83 +106549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143348,83 +106773,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143649,83 +106997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143951,83 +107222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144252,83 +107446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144554,83 +107671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144856,83 +107896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145157,83 +108120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145458,83 +108344,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145759,83 +108568,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146060,83 +108792,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146362,83 +109017,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146664,83 +109242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146966,83 +109467,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147267,83 +109691,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147569,83 +109916,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147870,83 +110140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148172,83 +110365,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148473,83 +110589,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148775,83 +110814,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149076,83 +111038,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149377,83 +111262,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149678,83 +111486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149979,83 +111710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150280,83 +111934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150582,83 +112159,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150884,83 +112384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151185,83 +112608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151487,83 +112833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151789,83 +113058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152090,83 +113282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152392,83 +113507,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152694,83 +113732,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152995,83 +113956,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153296,83 +114180,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153597,83 +114404,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153898,83 +114628,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154199,83 +114852,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154500,83 +115076,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154801,83 +115300,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155102,83 +115524,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155403,83 +115748,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155704,83 +115972,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156006,83 +116197,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156307,83 +116421,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156608,83 +116645,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156909,83 +116869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157210,83 +117093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157512,83 +117318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157813,83 +117542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158114,83 +117766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158415,83 +117990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158716,83 +118214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159017,83 +118438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159319,83 +118663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159620,83 +118887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159921,83 +119111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160223,83 +119336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160524,83 +119560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160826,83 +119785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161128,83 +120010,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161430,83 +120235,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161731,83 +120459,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162032,83 +120683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162333,83 +120907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162634,83 +121131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162936,83 +121356,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163237,83 +121580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163538,83 +121804,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163839,83 +122028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164140,83 +122252,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164441,83 +122476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164742,83 +122700,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165043,83 +122924,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165344,83 +123148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165645,83 +123372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165946,83 +123596,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166247,83 +123820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166549,83 +124045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166850,83 +124269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167151,83 +124493,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167452,83 +124717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167753,83 +124941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168054,83 +125165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168355,83 +125389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168656,83 +125613,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168957,83 +125837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169258,83 +126061,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169560,83 +126286,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169862,83 +126511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170163,83 +126735,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170465,83 +126960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170766,83 +127184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171067,83 +127408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171368,83 +127632,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171669,83 +127856,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171970,83 +128080,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172271,83 +128304,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172573,83 +128529,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172875,83 +128754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173176,83 +128978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173477,83 +129202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173779,83 +129427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174080,83 +129651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174381,83 +129875,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174682,83 +130099,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174983,83 +130323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175284,83 +130547,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175586,83 +130772,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175888,83 +130997,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176189,83 +131221,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176490,83 +131445,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176791,83 +131669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177093,83 +131894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177394,83 +132118,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177695,83 +132342,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177996,83 +132566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178297,83 +132790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178598,83 +133014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178899,83 +133238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179200,83 +133462,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179501,83 +133686,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179802,83 +133910,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180103,83 +134134,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180404,83 +134358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180705,83 +134582,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181007,83 +134807,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181308,83 +135031,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181609,83 +135255,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181910,83 +135479,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182211,83 +135703,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182512,83 +135927,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182813,83 +136151,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183114,83 +136375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183415,83 +136599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183717,83 +136824,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184018,83 +137048,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184319,83 +137272,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184620,83 +137496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184921,83 +137720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185222,83 +137944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185523,83 +138168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185824,83 +138392,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186126,83 +138617,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186428,83 +138842,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186729,83 +139066,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187031,83 +139291,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187332,83 +139515,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187634,83 +139740,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187936,83 +139965,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188238,83 +140190,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188539,83 +140414,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188840,83 +140638,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189141,83 +140862,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189442,83 +141086,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189743,83 +141310,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190045,83 +141535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190346,83 +141759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190648,83 +141984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190949,83 +142208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191250,83 +142432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191552,83 +142657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191854,83 +142882,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192156,83 +143107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192457,83 +143331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192759,83 +143556,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193060,83 +143780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193361,83 +144004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193663,83 +144229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193965,83 +144454,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194266,83 +144678,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194568,83 +144903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194869,83 +145127,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195170,83 +145351,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195471,83 +145575,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195773,83 +145800,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196074,83 +146024,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196376,83 +146249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196677,83 +146473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196978,83 +146697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197279,83 +146921,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197580,83 +147145,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197882,83 +147370,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198183,83 +147594,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198484,83 +147818,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198785,83 +148042,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199087,83 +148267,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199389,83 +148492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199691,83 +148717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199992,83 +148941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200294,83 +149166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200595,83 +149390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200896,83 +149614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201197,83 +149838,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201498,83 +150062,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201799,83 +150286,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202101,83 +150511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202403,83 +150736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202704,83 +150960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203006,83 +151185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203307,83 +151409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203609,83 +151634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203911,83 +151859,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204212,83 +152083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204513,83 +152307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204814,83 +152531,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205115,83 +152755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205417,83 +152980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205718,83 +153204,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206020,83 +153429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206321,83 +153653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206622,83 +153877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206924,83 +154102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207226,83 +154327,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207527,83 +154551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207828,83 +154775,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208130,83 +155000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208431,83 +155224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208732,83 +155448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209034,83 +155673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209335,83 +155897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209636,83 +156121,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209937,83 +156345,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210238,83 +156569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210539,83 +156793,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210841,83 +157018,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211142,83 +157242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211443,83 +157466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211744,83 +157690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212045,83 +157914,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212346,83 +158138,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212647,83 +158362,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212948,83 +158586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213250,83 +158811,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213551,83 +159035,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213852,83 +159259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214154,83 +159484,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214455,83 +159708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214757,83 +159933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215059,83 +160158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215360,83 +160382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215661,83 +160606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215962,83 +160830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216264,83 +161055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216566,83 +161280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216867,83 +161504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217169,83 +161729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217470,83 +161953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217772,83 +162178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218073,83 +162402,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218374,83 +162626,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218675,83 +162850,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218977,83 +163075,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219278,83 +163299,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219579,83 +163523,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219880,83 +163747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220181,83 +163971,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220482,83 +164195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220783,83 +164419,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221084,83 +164643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221385,83 +164867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221687,83 +165092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221988,83 +165316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222289,83 +165540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222591,83 +165765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222893,83 +165990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223194,83 +166214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223495,83 +166438,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223796,83 +166662,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224098,83 +166887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224400,83 +167112,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224701,83 +167336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225002,83 +167560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225304,83 +167785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225605,83 +168009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225907,83 +168234,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226208,83 +168458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226509,83 +168682,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226811,83 +168907,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227112,83 +169131,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227413,83 +169355,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227715,83 +169580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228016,83 +169804,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228317,83 +170028,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228619,83 +170253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228921,83 +170478,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229222,83 +170702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229523,83 +170926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229824,83 +171150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230125,83 +171374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230427,83 +171599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230728,83 +171823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231029,83 +172047,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231331,83 +172272,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231632,83 +172496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231933,83 +172720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232234,83 +172944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232535,83 +173168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232836,83 +173392,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233137,83 +173616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233438,83 +173840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233739,83 +174064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234041,83 +174289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234342,83 +174513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234643,83 +174737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234944,83 +174961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235245,83 +175185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235546,83 +175409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235847,83 +175633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236148,83 +175857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236449,83 +176081,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236750,83 +176305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237052,83 +176530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237354,83 +176755,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237656,83 +176980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237957,83 +177204,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238258,83 +177428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238559,83 +177652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238860,83 +177876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239162,83 +178101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239463,83 +178325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239764,83 +178549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240066,83 +178774,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240367,83 +178998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240668,83 +179222,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240969,83 +179446,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241270,83 +179670,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241572,83 +179895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241873,83 +180119,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242174,83 +180343,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242475,83 +180567,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242776,83 +180791,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243077,83 +181015,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243378,83 +181239,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243679,83 +181463,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243981,83 +181688,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244282,83 +181912,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244583,83 +182136,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244885,83 +182361,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245187,83 +182586,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245488,83 +182810,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245789,83 +183034,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246091,83 +183259,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246392,83 +183483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246694,83 +183708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246996,83 +183933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247298,83 +184158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247599,83 +184382,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247900,83 +184606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248201,83 +184830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248502,83 +185054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248803,83 +185278,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249104,83 +185502,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249405,83 +185726,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249706,83 +185950,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250007,83 +186174,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250308,83 +186398,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250609,83 +186622,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250910,83 +186846,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251212,83 +187071,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251513,83 +187295,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251815,83 +187520,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252116,83 +187744,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252417,83 +187968,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252718,83 +188192,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253019,83 +188416,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253320,83 +188640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253621,83 +188864,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253922,83 +189088,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254224,83 +189313,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254525,83 +189537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254826,83 +189761,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255127,83 +189985,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255428,83 +190209,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255730,83 +190434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256032,83 +190659,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256333,83 +190883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256635,83 +191108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256937,83 +191333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257239,83 +191558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257540,83 +191782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257841,83 +192006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258142,83 +192230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DataTypeA: 4 - DataTypeAmaxD: 0 - DataTypeB: 4 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258289,7 +192300,6 @@ reorderGRInstForDTVB: false tailLoopOptA: false tailLoopOptB: false - - [2, 3, 0, 1] - - - [16, 16, 1, 256] - [3, 0.0] diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8BH_AI_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8BH_AI_SAV_UserArgs.yaml index 749220c383e..7c2b6e2e31f 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8BH_AI_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8BH_AI_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -70,7 +70,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 UseScaleCD: false - - 1LDSBuffer: 1 @@ -206,78 +206,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -465,78 +393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -724,78 +580,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -983,78 +767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1242,78 +954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8I8S_BH_HAS_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8I8S_BH_HAS_Bias_HAS_SAV_UserArgs.yaml index 55618e291d6..127bc0c962b 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8I8S_BH_HAS_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8I8S_BH_HAS_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -222,83 +222,6 @@ PrefetchGlobalRead: 0 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 0 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8II_BH_AI_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8II_BH_AI_SAV_UserArgs.yaml index 14360de53a1..48d6b2351f7 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8II_BH_AI_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8II_BH_AI_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -70,7 +70,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 UseScaleCD: false - - 1LDSBuffer: 0 @@ -207,78 +207,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -461,77 +389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -713,77 +570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -965,77 +751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1217,77 +932,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1469,77 +1113,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeB: 8 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8II_BH_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8II_BH_UserArgs.yaml index 624e5911872..8f72889fea8 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8II_BH_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8II_BH_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -233,83 +233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -535,83 +458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -837,83 +683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1139,83 +908,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1441,83 +1133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1743,83 +1358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2045,83 +1583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2347,83 +1808,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2649,83 +2033,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2951,83 +2258,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3253,83 +2483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3555,83 +2708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3857,83 +2933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4159,83 +3158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4461,83 +3383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4763,83 +3608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5065,83 +3833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5367,83 +4058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5669,83 +4283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5971,83 +4508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6273,83 +4733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6575,83 +4958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6877,83 +5183,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7179,83 +5408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7481,83 +5633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7783,83 +5858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8085,83 +6083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8387,83 +6308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8689,83 +6533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8991,83 +6758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9293,83 +6983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9595,83 +7208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9897,83 +7433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10199,83 +7658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10501,83 +7883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10803,83 +8108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11105,83 +8333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11407,83 +8558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11709,83 +8783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12011,83 +9008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12313,83 +9233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12615,83 +9458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12917,83 +9683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13219,83 +9908,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13521,83 +10133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13823,83 +10358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14125,83 +10583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14427,83 +10808,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14729,83 +11033,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15031,83 +11258,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15333,83 +11483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15635,83 +11708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 2 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15937,83 +11933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 2 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16239,83 +12158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 2 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8IS_BH_HAS_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8IS_BH_HAS_Bias_HAS_SAV_UserArgs.yaml index e91f9b8ebc5..6e054cd9b32 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8IS_BH_HAS_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8IS_BH_HAS_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -222,83 +222,6 @@ PrefetchGlobalRead: 0 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 0 - DataTypeB: 8 - DataTypeE: 6 - DestDataType: 6 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8_BH_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8_BH_UserArgs.yaml index 14e9226d195..dd2b497c1fa 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8_BH_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_I8_BH_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -233,83 +233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -535,83 +458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -837,83 +683,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1139,83 +908,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1441,83 +1133,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1743,83 +1358,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2045,83 +1583,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2347,83 +1808,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 0 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2649,83 +2033,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2951,83 +2258,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3253,83 +2483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3555,83 +2708,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3857,83 +2933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4159,83 +3158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4461,83 +3383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4763,83 +3608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5065,83 +3833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5367,83 +4058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5669,83 +4283,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5971,83 +4508,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6273,83 +4733,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6575,83 +4958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6877,83 +5183,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7179,83 +5408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7481,83 +5633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7783,83 +5858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8085,83 +6083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8387,83 +6308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8689,83 +6533,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8991,83 +6758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9293,83 +6983,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9595,83 +7208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9897,83 +7433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10199,83 +7658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10501,83 +7883,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10803,83 +8108,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11105,83 +8333,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11407,83 +8558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 2 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11709,83 +8783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 2 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12011,83 +9008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 2 PreloadKernArgs: true - ProblemType: - Activation: false - ActivationComputeDataType: 6 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 6 - DataType: 8 - DataTypeA: 8 - DataTypeAmaxD: 6 - DataTypeB: 8 - DataTypeE: 8 - DestDataType: 8 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_Aux_A_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_Aux_A_SAV_UserArgs.yaml index 8bc98585e51..e881ca2c122 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_Aux_A_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_Aux_A_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -67,7 +67,7 @@ UseE: true UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 - - 1LDSBuffer: 1 ActivationAlt: false @@ -195,74 +195,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 1 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_BiasSrcB_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_BiasSrcB_SAV_UserArgs.yaml index 475230c2c94..f2d94f5aa86 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_BiasSrcB_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_BiasSrcB_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -67,7 +67,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 - - 1LDSBuffer: 1 ActivationAlt: false @@ -195,74 +195,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_BiasSrcD_Grad_A_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_BiasSrcD_Grad_A_SAV_UserArgs.yaml index 2c524748c45..59a9a0515b5 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_BiasSrcD_Grad_A_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_SB_Bias_BiasSrcD_Grad_A_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -67,7 +67,7 @@ UseE: true UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 1 - - 1LDSBuffer: 1 ActivationAlt: false @@ -195,74 +195,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_S_B_Bias_HAS_SAV_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_S_B_Bias_HAS_SAV_UserArgs.yaml index b77e07f1897..9f62a6f075c 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_S_B_Bias_HAS_SAV_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_S_B_Bias_HAS_SAV_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 75a0] @@ -232,83 +232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -533,83 +456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -835,83 +681,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1136,83 +905,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1437,83 +1129,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -1739,83 +1354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2040,83 +1578,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2342,83 +1803,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2643,83 +2027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -2944,83 +2251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3246,83 +2476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3547,83 +2700,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -3848,83 +2924,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4149,83 +3148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4450,83 +3372,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -4751,83 +3596,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5053,83 +3821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5355,83 +4046,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5656,83 +4270,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -5957,83 +4494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6259,83 +4719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6561,83 +4944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -6862,83 +5168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7163,83 +5392,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7464,83 +5616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -7765,83 +5840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8066,83 +6064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8368,83 +6289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8669,83 +6513,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -8970,83 +6737,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9271,83 +6961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9572,83 +7185,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -9874,83 +7410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10176,83 +7635,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10478,83 +7860,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -10779,83 +8084,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11080,83 +8308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11381,83 +8532,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11683,83 +8757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -11984,83 +8981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12285,83 +9205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12586,83 +9429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -12887,83 +9653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13188,83 +9877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13489,83 +10101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -13791,83 +10326,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14093,83 +10551,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14395,83 +10776,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14696,83 +11000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -14997,83 +11224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15298,83 +11448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15600,83 +11673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -15901,83 +11897,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16202,83 +12121,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16503,83 +12345,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -16804,83 +12569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17105,83 +12793,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17406,83 +13017,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -17708,83 +13242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18010,83 +13467,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18311,83 +13691,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18613,83 +13916,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -18914,83 +14140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19215,83 +14364,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19516,83 +14588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -19818,83 +14813,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20119,83 +15037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20420,83 +15261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -20721,83 +15485,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21022,83 +15709,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21323,83 +15933,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21624,83 +16157,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -21925,83 +16381,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22226,83 +16605,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22528,83 +16830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -22830,83 +17055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23132,83 +17280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23433,83 +17504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -23735,83 +17729,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24036,83 +17953,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24337,83 +18177,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24638,83 +18401,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -24939,83 +18625,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25240,83 +18849,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25541,83 +19073,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -25842,83 +19297,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26143,83 +19521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26445,83 +19746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -26746,83 +19970,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27048,83 +20195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27349,83 +20419,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27650,83 +20643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -27951,83 +20867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28252,83 +21091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28554,83 +21316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -28856,83 +21541,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29158,83 +21766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29459,83 +21990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -29761,83 +22215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30062,83 +22439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30363,83 +22663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30664,83 +22887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -30965,83 +23111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31267,83 +23336,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31568,83 +23560,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -31870,83 +23785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32171,83 +24009,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32472,83 +24233,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -32774,83 +24458,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33075,83 +24682,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33376,83 +24906,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33677,83 +25130,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -33978,83 +25354,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34279,83 +25578,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34580,83 +25802,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -34882,83 +26027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35184,83 +26252,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35485,83 +26476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -35786,83 +26700,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36087,83 +26924,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36388,83 +27148,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36690,83 +27373,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -36991,83 +27597,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37292,83 +27821,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37593,83 +28045,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -37894,83 +28269,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38196,83 +28494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38497,83 +28718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -38798,83 +28942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39100,83 +29167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39401,83 +29391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -39702,83 +29615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40003,83 +29839,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40304,83 +30063,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40605,83 +30287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -40906,83 +30511,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41207,83 +30735,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41509,83 +30960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -41810,83 +31184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42112,83 +31409,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42414,83 +31634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -42715,83 +31858,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43016,83 +32082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43317,83 +32306,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43618,83 +32530,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -43919,83 +32754,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44220,83 +32978,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44522,83 +33203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -44823,83 +33427,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45125,83 +33652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45426,83 +33876,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -45727,83 +34100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46028,83 +34324,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46329,83 +34548,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46630,83 +34772,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -46931,83 +34996,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47232,83 +35220,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47533,83 +35444,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -47834,83 +35668,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48135,83 +35892,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48436,83 +36116,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -48737,83 +36340,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49039,83 +36565,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49341,83 +36790,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49642,83 +37014,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -49943,83 +37238,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50244,83 +37462,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50546,83 +37687,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -50848,83 +37912,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51150,83 +38137,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51451,83 +38361,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -51752,83 +38585,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52053,83 +38809,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52355,83 +39034,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52656,83 +39258,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -52958,83 +39483,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53259,83 +39707,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53560,83 +39931,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -53861,83 +40155,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54163,83 +40380,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54464,83 +40604,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -54765,83 +40828,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55066,83 +41052,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55367,83 +41276,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55668,83 +41500,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -55969,83 +41724,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56270,83 +41948,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56572,83 +42173,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -56873,83 +42397,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57175,83 +42622,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57476,83 +42846,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -57777,83 +43070,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58078,83 +43294,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58379,83 +43518,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58680,83 +43742,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -58982,83 +43967,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59283,83 +44191,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59584,83 +44415,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -59885,83 +44639,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60186,83 +44863,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60488,83 +45088,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -60789,83 +45312,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61090,83 +45536,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61391,83 +45760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61692,83 +45984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -61993,83 +46208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62295,83 +46433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62596,83 +46657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -62897,83 +46881,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63198,83 +47105,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63500,83 +47330,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -63801,83 +47554,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64103,83 +47779,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64405,83 +48004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -64706,83 +48228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65008,83 +48453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65310,83 +48678,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65612,83 +48903,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -65914,83 +49128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66215,83 +49352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66517,83 +49577,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -66819,83 +49802,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67121,83 +50027,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67422,83 +50251,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -67724,83 +50476,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68026,83 +50701,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68328,83 +50926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68629,83 +51150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -68931,83 +51375,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69232,83 +51599,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69533,83 +51823,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -69834,83 +52047,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70135,83 +52271,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70436,83 +52495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -70738,83 +52720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71039,83 +52944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71340,83 +53168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71641,83 +53392,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -71942,83 +53616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72244,83 +53841,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72545,83 +54065,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -72846,83 +54289,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73148,83 +54514,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73449,83 +54738,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -73751,83 +54963,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74052,83 +55187,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74353,83 +55411,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74654,83 +55635,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -74955,83 +55859,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75256,83 +56083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75557,83 +56307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -75859,83 +56532,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76161,83 +56757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76462,83 +56981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -76764,83 +57206,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77065,83 +57430,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77366,83 +57654,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77668,83 +57879,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -77969,83 +58103,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78271,83 +58328,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78572,83 +58552,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -78873,83 +58776,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79174,83 +59000,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79475,83 +59224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -79776,83 +59448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80077,83 +59672,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80378,83 +59896,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80679,83 +60120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -80981,83 +60345,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81282,83 +60569,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81583,83 +60793,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -81884,83 +61017,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82185,83 +61241,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82487,83 +61466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -82788,83 +61690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83090,83 +61915,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83392,83 +62140,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83693,83 +62364,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -83994,83 +62588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84295,83 +62812,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84597,83 +63037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -84899,83 +63262,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85201,83 +63487,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85503,83 +63712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -85804,83 +63936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86106,83 +64161,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86408,83 +64386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -86710,83 +64611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87012,83 +64836,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87313,83 +65060,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87615,83 +65285,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -87917,83 +65510,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88218,83 +65734,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88519,83 +65958,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -88820,83 +66182,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89121,83 +66406,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89422,83 +66630,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -89723,83 +66854,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90024,83 +67078,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90326,83 +67303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90627,83 +67527,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -90928,83 +67751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91229,83 +67975,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91530,83 +68199,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -91831,83 +68423,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92132,83 +68647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92433,83 +68871,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -92734,83 +69095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93035,83 +69319,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93336,83 +69543,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93637,83 +69767,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -93938,83 +69991,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94239,83 +70215,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94540,83 +70439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -94841,83 +70663,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95142,83 +70887,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95443,83 +71111,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -95744,83 +71335,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96045,83 +71559,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96346,83 +71783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96647,83 +72007,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -96949,83 +72232,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97250,83 +72456,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97551,83 +72680,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -97852,83 +72904,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98153,83 +73128,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98454,83 +73352,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -98755,83 +73576,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99057,83 +73801,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99358,83 +74025,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99659,83 +74249,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -99960,83 +74473,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100261,83 +74697,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100563,83 +74922,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -100864,83 +75146,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101166,83 +75371,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101467,83 +75595,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -101769,83 +75820,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102070,83 +76044,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102371,83 +76268,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102672,83 +76492,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -102973,83 +76716,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103274,83 +76940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103576,83 +77165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -103877,83 +77389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104178,83 +77613,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104480,83 +77838,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -104782,83 +78063,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105083,83 +78287,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105385,83 +78512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105686,83 +78736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -105987,83 +78960,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106288,83 +79184,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106589,83 +79408,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -106891,83 +79633,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107193,83 +79858,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107494,83 +80082,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -107796,83 +80307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108098,83 +80532,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108400,83 +80757,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -108701,83 +80981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109002,83 +81205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109303,83 +81429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109604,83 +81653,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -109905,83 +81877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110206,83 +82101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110507,83 +82325,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -110808,83 +82549,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111110,83 +82774,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111411,83 +82998,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -111713,83 +83223,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112015,83 +83448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112317,83 +83673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112619,83 +83898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -112920,83 +84122,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113221,83 +84346,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113522,83 +84570,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -113823,83 +84794,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114124,83 +85018,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114425,83 +85242,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -114726,83 +85466,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115027,83 +85690,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115328,83 +85914,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115629,83 +86138,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -115931,83 +86363,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116233,83 +86588,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116535,83 +86813,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -116836,83 +87037,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117137,83 +87261,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117439,83 +87486,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -117740,83 +87710,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118041,83 +87934,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118342,83 +88158,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118644,83 +88383,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -118945,83 +88607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119246,83 +88831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119547,83 +89055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -119848,83 +89279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120149,83 +89503,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120450,83 +89727,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -120751,83 +89951,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121052,83 +90175,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121353,83 +90399,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121655,83 +90624,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -121956,83 +90848,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122257,83 +91072,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122558,83 +91296,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -122860,83 +91521,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123162,83 +91746,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123463,83 +91970,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -123764,83 +92194,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124065,83 +92418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124367,83 +92643,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124668,83 +92867,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -124969,83 +93091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125271,83 +93316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125573,83 +93541,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -125874,83 +93765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126175,83 +93989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126477,83 +94214,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -126779,83 +94439,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127081,83 +94664,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127383,83 +94889,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127685,83 +95114,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -127986,83 +95338,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128288,83 +95563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128589,83 +95787,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -128891,83 +96012,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129192,83 +96236,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129493,83 +96460,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -129794,83 +96684,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130095,83 +96908,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130396,83 +97132,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130697,83 +97356,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -130999,83 +97581,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131300,83 +97805,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131601,83 +98029,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -131902,83 +98253,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132204,83 +98478,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132505,83 +98702,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -132806,83 +98926,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133107,83 +99150,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133408,83 +99374,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -133709,83 +99598,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134010,83 +99822,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134312,83 +100047,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134613,83 +100271,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -134914,83 +100495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135215,83 +100719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135514,83 +100941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -135813,83 +101163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136112,83 +101385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136411,83 +101607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -136710,83 +101829,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137009,83 +102051,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137308,83 +102273,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137607,83 +102495,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -137906,83 +102717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138205,83 +102939,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138505,83 +103162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -138804,83 +103384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139103,83 +103606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139402,83 +103828,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -139701,83 +104050,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140000,83 +104272,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140299,83 +104494,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140599,83 +104717,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -140898,83 +104939,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141198,83 +105162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141497,83 +105384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -141797,83 +105607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142097,83 +105830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142396,83 +106052,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142695,83 +106274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -142994,83 +106496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143293,83 +106718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143593,83 +106941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -143892,83 +107163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144191,83 +107385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144490,83 +107607,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -144790,83 +107830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145089,83 +108052,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145388,83 +108274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145688,83 +108497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -145987,83 +108719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146286,83 +108941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146586,83 +109164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -146886,83 +109387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147185,83 +109609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147485,83 +109832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -147784,83 +110054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148084,83 +110277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148383,83 +110499,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148682,83 +110721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -148981,83 +110943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149280,83 +111165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149579,83 +111387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -149878,83 +111609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150178,83 +111832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150478,83 +112055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -150777,83 +112277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151076,83 +112499,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151375,83 +112721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151674,83 +112943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -151973,83 +113165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152272,83 +113387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152571,83 +113609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -152870,83 +113831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153169,83 +114053,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153469,83 +114276,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -153768,83 +114498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154067,83 +114720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154366,83 +114942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154666,83 +115165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -154965,83 +115387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155264,83 +115609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155563,83 +115831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -155862,83 +116053,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156161,83 +116275,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156460,83 +116497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -156759,83 +116719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157058,83 +116941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157358,83 +117164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157657,83 +117386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -157956,83 +117608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158255,83 +117830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158554,83 +118052,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -158853,83 +118274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159152,83 +118496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159451,83 +118718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -159750,83 +118940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160049,83 +119162,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160348,83 +119384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160647,83 +119606,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -160947,83 +119829,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161246,83 +120051,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161546,83 +120274,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -161845,83 +120496,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162144,83 +120718,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162443,83 +120940,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -162743,83 +121163,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163042,83 +121385,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163342,83 +121608,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163641,83 +121830,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -163941,83 +122053,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164240,83 +122275,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164539,83 +122497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -164838,83 +122719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165138,83 +122942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165438,83 +123165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -165737,83 +123387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166036,83 +123609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166335,83 +123831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166634,83 +124053,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -166934,83 +124276,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167233,83 +124498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167532,83 +124720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -167831,83 +124942,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168131,83 +125165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168430,83 +125387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -168730,83 +125610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169029,83 +125832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169329,83 +126055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169628,83 +126277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -169927,83 +126499,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170226,83 +126721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170525,83 +126943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -170824,83 +127165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171123,83 +127387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171423,83 +127610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -171722,83 +127832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172021,83 +128054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172320,83 +128276,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172619,83 +128498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -172919,83 +128721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173218,83 +128943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173517,83 +129165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -173816,83 +129387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174115,83 +129609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174414,83 +129831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -174713,83 +130053,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175012,83 +130275,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175311,83 +130497,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175610,83 +130719,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -175909,83 +130941,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176209,83 +131164,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176508,83 +131386,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -176808,83 +131609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177107,83 +131831,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177407,83 +132054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -177706,83 +132276,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178006,83 +132499,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178305,83 +132721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178604,83 +132943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -178903,83 +133165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179202,83 +133387,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179502,83 +133610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -179801,83 +133832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180100,83 +134054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180399,83 +134276,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180698,83 +134498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -180998,83 +134721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181297,83 +134943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181597,83 +135166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -181896,83 +135388,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182195,83 +135610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182494,83 +135832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -182793,83 +136054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183092,83 +136276,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183391,83 +136498,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183690,83 +136720,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -183990,83 +136943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184289,83 +137165,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184589,83 +137388,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -184889,83 +137611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185188,83 +137833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185487,83 +138055,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -185786,83 +138277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186085,83 +138499,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186384,83 +138721,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186683,83 +138943,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -186983,83 +139166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187282,83 +139388,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187581,83 +139610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -187880,83 +139832,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188179,83 +140054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188479,83 +140277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -188779,83 +140500,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189078,83 +140722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189378,83 +140945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189677,83 +141167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -189976,83 +141389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190275,83 +141611,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190574,83 +141833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -190873,83 +142055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191172,83 +142277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191471,83 +142499,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -191771,83 +142722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192070,83 +142944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192369,83 +143166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192668,83 +143388,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -192967,83 +143610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193266,83 +143832,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193565,83 +144054,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -193865,83 +144277,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194164,83 +144499,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194464,83 +144722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -194764,83 +144945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195063,83 +145167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195363,83 +145390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195662,83 +145612,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -195962,83 +145835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196261,83 +146057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196561,83 +146280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -196860,83 +146502,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197159,83 +146724,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197458,83 +146946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -197757,83 +147168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198056,83 +147390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198355,83 +147612,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198654,83 +147834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -198953,83 +148056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199252,83 +148278,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199551,83 +148500,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -199851,83 +148723,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200151,83 +148946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200450,83 +149168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -200749,83 +149390,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201048,83 +149612,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201347,83 +149834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201646,83 +150056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -201945,83 +150278,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202244,83 +150500,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202543,83 +150722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -202842,83 +150944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203141,83 +151166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203440,83 +151388,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -203739,83 +151610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204039,83 +151833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204339,83 +152056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204638,83 +152278,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -204937,83 +152500,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205236,83 +152722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205535,83 +152944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -205834,83 +153166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206133,83 +153388,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206432,83 +153610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -206732,83 +153833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207032,83 +154056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207331,83 +154278,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207630,83 +154500,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -207929,83 +154722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208228,83 +154944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208528,83 +155167,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -208827,83 +155389,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209127,83 +155612,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209426,83 +155834,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -209725,83 +156056,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210024,83 +156278,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210323,83 +156500,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210622,83 +156722,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -210921,83 +156944,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211220,83 +157166,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211519,83 +157388,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -211818,83 +157610,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212118,83 +157833,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212417,83 +158055,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -212717,83 +158278,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213017,83 +158501,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213317,83 +158724,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213616,83 +158946,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -213915,83 +159168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214215,83 +159391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214514,83 +159613,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -214813,83 +159835,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215113,83 +160058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215412,83 +160280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -215711,83 +160502,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216011,83 +160725,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216310,83 +160947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216610,83 +161170,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -216910,83 +161393,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217209,83 +161615,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217508,83 +161837,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -217807,83 +162059,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218107,83 +162282,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218406,83 +162504,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -218705,83 +162726,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219004,83 +162948,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219303,83 +163170,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219602,83 +163392,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -219901,83 +163614,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220200,83 +163836,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220499,83 +164058,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -220798,83 +164280,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221097,83 +164502,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221396,83 +164724,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221696,83 +164947,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -221995,83 +165169,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222294,83 +165391,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222593,83 +165613,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -222892,83 +165835,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223191,83 +166057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223490,83 +166279,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -223789,83 +166501,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224088,83 +166723,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224387,83 +166945,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224687,83 +167168,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -224988,83 +167392,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225289,83 +167616,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225590,83 +167840,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -225891,83 +168064,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226192,83 +168288,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226493,83 +168512,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -226794,83 +168736,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227096,83 +168961,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227398,83 +169186,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -227699,83 +169410,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228000,83 +169634,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228302,83 +169859,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228603,83 +170083,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -228904,83 +170307,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229203,83 +170529,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229502,83 +170751,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -229801,83 +170973,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230100,83 +171195,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230400,83 +171418,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230699,83 +171640,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -230998,83 +171862,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231297,83 +172084,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231598,83 +172308,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -231899,83 +172532,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232200,83 +172756,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232500,83 +172979,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -232800,83 +173202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233100,83 +173425,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233399,83 +173647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233699,83 +173870,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -233998,83 +174092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234297,83 +174314,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234597,83 +174537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -234897,83 +174760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235196,83 +174982,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235495,83 +175204,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -235794,83 +175426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236095,83 +175650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236396,83 +175874,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236697,83 +176098,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -236996,83 +176320,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237295,83 +176542,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237594,83 +176764,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -237893,83 +176986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238192,83 +177208,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238491,83 +177430,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -238790,83 +177652,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239089,83 +177874,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239388,83 +178096,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239687,83 +178318,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -239986,83 +178540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240285,83 +178762,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240585,83 +178985,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -240884,83 +179207,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241183,83 +179429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241482,83 +179651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -241781,83 +179873,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242081,83 +180096,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242380,83 +180318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242679,83 +180540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -242978,83 +180762,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243277,83 +180984,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243576,83 +181206,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -243876,83 +181429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244175,83 +181651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244474,83 +181873,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -244773,83 +182095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245073,83 +182318,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245372,83 +182540,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245672,83 +182763,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -245971,83 +182985,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246270,83 +183207,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246569,83 +183429,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -246868,83 +183651,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247167,83 +183873,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247466,83 +184095,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -247765,83 +184317,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248064,83 +184539,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248363,83 +184761,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248665,83 +184986,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -248966,83 +185210,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249267,83 +185434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249568,83 +185658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -249869,83 +185882,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250170,83 +186106,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250472,83 +186331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -250773,83 +186555,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251075,83 +186780,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251376,83 +187004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251677,83 +187228,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -251979,83 +187453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252280,83 +187677,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252581,83 +187901,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -252882,83 +188125,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253183,83 +188349,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253484,83 +188573,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -253785,83 +188797,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254087,83 +189022,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254389,83 +189247,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254690,83 +189471,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -254992,83 +189696,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255293,83 +189920,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255594,83 +190144,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -255895,83 +190368,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256196,83 +190592,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256497,83 +190816,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -256798,83 +191040,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257099,83 +191264,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257400,83 +191488,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -257701,83 +191712,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258002,83 +191936,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258303,83 +192160,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258604,83 +192384,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -258906,83 +192609,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -259207,83 +192833,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -259508,83 +193057,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -259809,83 +193281,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -260110,83 +193505,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -260412,83 +193730,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -260713,83 +193954,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -261014,83 +194178,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -261316,83 +194403,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -261618,83 +194628,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -261920,83 +194853,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -262222,83 +195078,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -262524,83 +195303,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -262823,83 +195525,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -263122,83 +195747,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -263421,83 +195969,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -263720,83 +196191,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -264019,83 +196413,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -264318,83 +196635,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -264617,83 +196857,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -264918,83 +197081,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -265219,83 +197305,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -265520,83 +197529,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -265821,83 +197753,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -266122,83 +197977,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -266424,83 +198202,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -266723,83 +198424,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -267022,83 +198646,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -267322,83 +198869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -267621,83 +199091,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -267920,83 +199313,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -268219,83 +199535,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -268519,83 +199758,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -268818,83 +199980,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -269118,83 +200203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -269418,83 +200426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -269717,83 +200648,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -270016,83 +200870,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -270316,83 +201093,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -270615,83 +201315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -270914,83 +201537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -271213,83 +201759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -271512,83 +201981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -271811,83 +202203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -272111,83 +202426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -272410,83 +202648,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -272709,83 +202870,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -273008,83 +203092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -273308,83 +203315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -273607,83 +203537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -273906,83 +203759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -274205,83 +203981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -274504,83 +204203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -274804,83 +204426,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -275103,83 +204648,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -275402,83 +204870,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -275701,83 +205092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -276001,83 +205315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -276300,83 +205537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -276599,83 +205759,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -276898,83 +205981,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -277197,83 +206203,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -277496,83 +206425,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -277795,83 +206647,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -278094,83 +206869,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -278394,83 +207092,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -278694,83 +207315,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -278993,83 +207537,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -279293,83 +207760,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -279592,83 +207982,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -279892,83 +208205,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -280192,83 +208428,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -280491,83 +208650,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -280790,83 +208872,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -281089,83 +209094,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -281388,83 +209316,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -281690,83 +209541,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -281991,83 +209765,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -282292,83 +209989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -282592,83 +210212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -282891,83 +210434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -283190,83 +210656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -283490,83 +210879,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -283789,83 +211101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -284088,83 +211323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -284387,83 +211545,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -284687,83 +211768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -284986,83 +211990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -285285,83 +212212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -285584,83 +212434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -285883,83 +212656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -286182,83 +212878,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -286481,83 +213100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -286780,83 +213322,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -287079,83 +213544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -287378,83 +213766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -287678,83 +213989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -287977,83 +214211,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -288276,83 +214433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -288576,83 +214656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -288875,83 +214878,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -289175,83 +215101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -289475,83 +215324,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -289774,83 +215546,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -290073,83 +215768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -290372,83 +215990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -290671,83 +216212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -290970,83 +216434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -291270,83 +216657,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -291570,83 +216880,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -291869,83 +217102,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -292168,83 +217324,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -292467,83 +217546,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -292766,83 +217768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -293065,83 +217990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -293364,83 +218212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -293663,83 +218434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -293962,83 +218656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -294261,83 +218878,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -294560,83 +219100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -294859,83 +219322,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -295158,83 +219544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -295457,83 +219766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -295756,83 +219988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -296055,83 +220210,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -296354,83 +220432,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -296654,83 +220655,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -296953,83 +220877,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -297252,83 +221099,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -297551,83 +221321,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -297851,83 +221544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -298150,83 +221766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -298449,83 +221988,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -298748,83 +222210,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -299048,83 +222433,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -299348,83 +222656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -299647,83 +222878,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -299947,83 +223101,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -300246,83 +223323,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -300545,83 +223545,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -300845,83 +223768,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -301144,83 +223990,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -301443,83 +224212,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -301742,83 +224434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -302041,83 +224656,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -302340,83 +224878,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -302639,83 +225100,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -302938,83 +225322,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -303237,83 +225544,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -303536,83 +225766,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -303836,83 +225989,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -304135,83 +226211,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -304435,83 +226434,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -304736,83 +226658,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -305037,83 +226882,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -305339,83 +227107,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -305640,83 +227331,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -305941,83 +227555,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -306242,83 +227779,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -306542,83 +228002,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -306841,83 +228224,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -307141,83 +228447,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -307440,83 +228669,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -307739,83 +228891,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -308038,83 +229113,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -308337,83 +229335,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -308637,83 +229558,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -308937,83 +229781,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -309237,83 +230004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -309536,83 +230226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -309835,83 +230448,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -310135,83 +230671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -310434,83 +230893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -310733,83 +231115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -311032,83 +231337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -311331,83 +231559,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -311631,83 +231782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -311931,83 +232005,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -312230,83 +232227,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -312529,83 +232449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -312828,83 +232671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -313127,83 +232893,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -313426,83 +233115,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -313725,83 +233337,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -314024,83 +233559,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -314324,83 +233782,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -314623,83 +234004,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -314922,83 +234226,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -315222,83 +234449,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -315521,83 +234671,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -315821,83 +234894,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -316120,83 +235116,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -316420,83 +235339,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -316719,83 +235561,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -317018,83 +235783,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -317318,83 +236006,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -317618,83 +236229,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -317917,83 +236451,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -318216,83 +236673,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -318515,83 +236895,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -318814,83 +237117,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -319114,83 +237340,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -319414,83 +237563,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -319713,83 +237785,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -320013,83 +238008,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -320312,83 +238230,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -320612,83 +238453,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -320912,83 +238676,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -321211,83 +238898,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -321510,83 +239120,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -321810,83 +239343,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -322110,83 +239566,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -322410,83 +239789,6 @@ PrefetchGlobalRead: 2 PrefetchLocalRead: 1 PreloadKernArgs: true - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: true - BiasDataTypeList: [0] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: false - TransposeB: false - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 1 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_S_MX_B_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_S_MX_B_UserArgs.yaml index 2d2221d8fca..34b9c8aab71 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_S_MX_B_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Ailk_Bljk_S_MX_B_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049, Device 0050] @@ -227,83 +227,6 @@ PrefetchGlobalRead: 1 PrefetchLocalRead: 0 PreloadKernArgs: 0 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 0 - DataTypeA: 0 - DataTypeAmaxD: 0 - DataTypeB: 0 - DataTypeE: 0 - DestDataType: 0 - F32XdlMathOp: 10 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [0, 3, 2] - IndexAssignmentsB: [3, 1, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 1 - IndexUnrollB: 0 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - OutputAmaxD: false - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StochasticRounding: false - StridedBatched: true - SupportUserArgs: true - SwizzleTensorA: false - SwizzleTensorB: false - TLUA: true - TLUB: false - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 0 - TransposeB: 0 - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: '' - UseScaleAlphaVec: 0 - UseScaleCD: false ScheduleGlobalRead: 1 ScheduleIterAlg: 1 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_DB_UserArgs.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_DB_UserArgs.yaml index f2da34e2f10..26cf1e6ea09 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_DB_UserArgs.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_DB_UserArgs.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0058] @@ -67,7 +67,7 @@ UseE: false UseInitialStridesAB: false UseInitialStridesCD: false - UseScaleAB: "" + UseScaleAB: '' UseScaleAlphaVec: 0 - - 1LDSBuffer: 0 ActivationAlt: false @@ -194,74 +194,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 1 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 1 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - DataType: 1 - DestDataType: 1 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: true - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 0 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -440,74 +372,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 1 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 1 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 1 - DataType: 1 - DestDataType: 1 - F32XdlMathOp: 0 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: false - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexAssignmentsMetadata: [3, 0, 2] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndexUnrollM: 0 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - MirrorDimsMetadata: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SetConstStrideBias: [] - SilentHighPrecisionAccumulate: false - Sparse: 0 - StridedBatched: true - SupportUserArgs: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: true - TransposeB: true - UseBeta: true - UseBias: 0 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAB: "" - UseScaleAlphaVec: 0 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_Aux_AH_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_Aux_AH_SAV.yaml index 554e19fa47e..670bd115424 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_Aux_AH_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_Aux_AH_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -401,66 +341,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -625,66 +505,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: false - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcA_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcA_SAV.yaml index 86477ac44e4..de43f48fa71 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcA_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcA_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: A - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -400,66 +340,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: A - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -623,66 +503,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: A - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcB_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcB_SAV.yaml index f60b385e6ac..9ce0bc56d6f 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcB_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcB_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -400,66 +340,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: false - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: none - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: B - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: false - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml index 0cc536d5c5a..8c72593a112 100644 --- a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml +++ b/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBased/gfx950_Cijk_Alik_Bjlk_HHS_BH_Bias_BiasSrcD_Grad_AH_SAV.yaml @@ -1,4 +1,4 @@ -- {MinimumRequiredVersion: 4.33.0} +- {MinimumRequiredVersion: 5.0.0} - gfx950 - gfx950 - [Device 0049] @@ -177,66 +177,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 @@ -401,66 +341,6 @@ PackedC1IndicesX: [1] PrefetchGlobalRead: 2 PrefetchLocalRead: 1 - ProblemType: - Activation: true - ActivationComputeDataType: 0 - ActivationNoGuard: false - ActivationType: hipblaslt_all - AllowNoFreeDims: false - AssignedDerivedParameters: true - Batched: true - BetaOnlyUseBias: false - BiasDataTypeList: [0, 4] - BiasSrc: D - ComplexConjugateA: false - ComplexConjugateB: false - ComputeDataType: 0 - DataType: 4 - DestDataType: 4 - Gradient: true - GroupedGemm: false - HighPrecisionAccumulate: true - Index0: 0 - Index01A: 0 - Index01B: 1 - Index1: 1 - IndexAssignmentsA: [3, 0, 2] - IndexAssignmentsB: [1, 3, 2] - IndexAssignmentsLD: [4, 5, 6, 7] - IndexUnroll: 3 - IndexUnrollA: 0 - IndexUnrollB: 1 - IndicesBatch: [2] - IndicesFree: [0, 1] - IndicesSummation: [3] - MirrorDimsA: [] - MirrorDimsB: [] - NumIndicesBatch: 1 - NumIndicesC: 3 - NumIndicesFree: 2 - NumIndicesLD: 4 - NumIndicesSummation: 1 - OperationType: GEMM - SetConstStrideA: [] - SetConstStrideB: [] - SilentHighPrecisionAccumulate: false - StridedBatched: true - TLUA: false - TLUB: true - Tensor0: 0 - Tensor1: 1 - TileA: 0 - TileAwareSelection: false - TileB: 1 - TotalIndices: 4 - TransposeA: 1 - TransposeB: 1 - UseBeta: true - UseBias: 1 - UseE: true - UseInitialStridesAB: false - UseInitialStridesCD: false - UseScaleAlphaVec: 1 ScheduleGlobalRead: 1 ScheduleIterAlg: 3 ScheduleLocalWrite: 1 diff --git a/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/gfx950/GridBa{"code":"deadline_exceeded","msg":"operation timed out"}